Models · 60 tracked · 48 sizeable · target Interactive from the Plan "My stack (Plan)" · recomputed in 7778 ms

43 of 60 tracked models run on a machine you can buy — 37 of them at ≥ 15 tok/s, and 6 pipelines on memory and a supported backend, at 32K context, out of 89 machines the search could build.

43 RUNNABLENO CONFIDENCE LEVEL

48 of 60 tracked models have both a shape and a measured quantization, so only those 48 carry a memory figure. The other 12 are counted, named and shown with what onboarding still owes — they are excluded from every coverage number on this screen, which is therefore a lower bound on what this catalog can run, not an estimate of it. Machines come from the same search the Plan and Compare use: 89 complete rigs built from 171 catalog parts.

Since yesterdayNo model was added, re-read from source or rewritten.compared with the model catalog as it stood Oct 20, 3:25 AM ET (24 h ago)

What moved

All 60 tracked models were already on record 24 h ago.

What this screen is showing

every role · 32K context

Every tracked model

60 of 60 rows · memory at 32K context · Qwen3-VL 235B-A22B Instruct is shown in full below

60 of 60 tracked models, with the memory each needs at 32K context and how many machines in the catalog run it
Model	Params B	Quantization	Needs GB	Machines	Onboarding	Confirmed at source
Gemma 4 E4B ITGoogle · dense · vision · 128K ctx max	7.5 (derived) DERIVEDGemma 4 E4B IT total parametersArithmetic on observed inputs only. No tunable assumption enters it. Σ params of language-model tensors — Cross-checked: tensor count 8B vs Hugging Face 8B (1.16e-5%). 7.5 B, counted from the tensor headers of https://huggingface.co/google/gemma-4-E4B-it at 2026-10-17T13:01:23.544Z total_parameters 7.5 B — https://huggingface.co/google/gemma-4-E4B-it counted_at 2026-10-17T13:01:23.544Z — spec_sources.verified_at for total_params_b — when this figure was counted; a refresh that keeps it does not rewrite it confidence 0.95 — spec_sources.confidence — from the tensor headers The parameters of the language-model tensors only, counted from the tensor headers: a vision or audio tower and multi-token-prediction heads are excluded, so it can be below the whole-checkpoint count the formula's cross-check quotes. It sizes the weights only when no measured file size is stored for the chosen quantization; it is not what is read per token on a mixture-of-experts model. Shown to one decimal: the underlying tensor count is exact, but more digits would claim a precision the published figure does not have (SPEC §14). 6.8 (derived) DERIVEDGemma 4 E4B IT active parameters per tokenArithmetic on observed inputs only. No tunable assumption enters it. active − input embedding table — Per-token memory read used for decode bandwidth (SPEC §8.2). 6.8 B, counted from the tensor headers of https://huggingface.co/google/gemma-4-E4B-it at 2026-10-17T13:01:23.544Z active_parameters_per_token 6.8 B — https://huggingface.co/google/gemma-4-E4B-it counted_at 2026-10-17T13:01:23.544Z — spec_sources.verified_at for active_params_b — when this figure was counted; a refresh that keeps it does not rewrite it confidence 0.95 — spec_sources.confidence — from the tensor headers The parameters read to produce one token. It sets the decode speed; on a dense model it equals the total. Shown to one decimal: the underlying tensor count is exact, but more digits would claim a precision the published figure does not have (SPEC §14). active	Q3_K_M 4.06 GB (observed) OBSERVEDQ3_K_M file size on diskMeasured or published, passed through unchanged — not computed here. quantization file size — published at the source, passed through unchanged; nothing is computed here 4.06 GB, as recorded from https://huggingface.co/unsloth/gemma-4-E4B-it-GGUF/tree/main, file gemma-4-E4B-it-Q3_K_M.gguf quantization_file_size 4.06 GB — https://huggingface.co/unsloth/gemma-4-E4B-it-GGUF/tree/main, file gemma-4-E4B-it-Q3_K_M.gguf verified_at 2026-10-17T13:01:23.544Z — the row's own verification timestamp quantization Q3_K_M — gguf bits_per_weight 4.3183 bits — the quantizer's own naming, or measured from the file The on-disk size of the weights. It is not the memory the model occupies: add the KV cache and the runtime overhead (SPEC §8.1). Preferred over the bits-per-weight formula wherever it exists, because it includes embeddings and any higher-precision tensors. Observed, not derived: this app records the figure and its source, and does not re-measure it. on disk	5.36 GB (estimate) ESTIMATEGemma 4 E4B IT memory requirement at 32K contextArithmetic that also depends on a tunable assumption, listed below. required_gb = weights_gb + kv_cache_gb + overhead_gb required_gb = 4.06 + 0.98 + 0.32 = 5.36 GB weights_gb 4.058137728 GB — real file size kv_cache_gb 0.976224256 GB — kv formula overhead_gb 0.32465101824000003 GB — overhead fraction Real file size used instead of the params × bits formula (SPEC §8.1). 7 layers keep a full KV cache, 35 sliding-window layers are capped at 512 tokens. Runtime buffers, activations, and allocator slack approximated as a fraction of weights.	89 of 89 (derived) DERIVEDmachines that run Gemma 4 E4B IT at ≥ 15 tok/sArithmetic on observed inputs only. No tunable assumption enters it. machines = #{candidate rigs where the model loads AND clears the Plan's speed floor} 89 evaluated → 89 hold it in memory → 89 of those also reach ≥ 15 tok/s candidates_evaluated 89 machines — every complete rig the search could build from the priced catalog hold_it_in_memory 89 machines — required memory ≤ usable memory (SPEC §8.6) clear_the_speed_floor 89 machines — ≥ 15 tok/s The denominator is the machines this catalog can build today, not every machine that exists. A model with no shape or no sized quantization is counted on no machine at all, and says so rather than counting as zero.	SIZEABLE	read stale 3 d — past the 36 hour freshness window — Gemma 4 E4B IT: the least recent read of its data from source (shape counted Oct 17, 9:01 AM ET from the tensor headers of https://huggingface.co/google/gemma-4-E4B-it; quantization sizes read Oct 17, 9:01 AM ET); never checked upstream.
Qwen3-VL 2B InstructAlibaba Qwen · dense · vision · 256K ctx max	1.7 (derived) DERIVEDQwen3-VL 2B Instruct total parametersArithmetic on observed inputs only. No tunable assumption enters it. Σ params of language-model tensors — Cross-checked: tensor count 2.13B vs Hugging Face 2.13B (0%). 1.7 B, counted from the tensor headers of https://huggingface.co/Qwen/Qwen3-VL-2B-Instruct at 2026-10-17T11:56:11.486Z total_parameters 1.7 B — https://huggingface.co/Qwen/Qwen3-VL-2B-Instruct counted_at 2026-10-17T11:56:11.486Z — spec_sources.verified_at for total_params_b — when this figure was counted; a refresh that keeps it does not rewrite it confidence 0.95 — spec_sources.confidence — from the tensor headers The parameters of the language-model tensors only, counted from the tensor headers: a vision or audio tower and multi-token-prediction heads are excluded, so it can be below the whole-checkpoint count the formula's cross-check quotes. It sizes the weights only when no measured file size is stored for the chosen quantization; it is not what is read per token on a mixture-of-experts model. Shown to one decimal: the underlying tensor count is exact, but more digits would claim a precision the published figure does not have (SPEC §14). 1.4 (derived) DERIVEDQwen3-VL 2B Instruct active parameters per tokenArithmetic on observed inputs only. No tunable assumption enters it. active − input embedding table — Per-token memory read used for decode bandwidth (SPEC §8.2). 1.4 B, counted from the tensor headers of https://huggingface.co/Qwen/Qwen3-VL-2B-Instruct at 2026-10-17T11:56:11.486Z active_parameters_per_token 1.4 B — https://huggingface.co/Qwen/Qwen3-VL-2B-Instruct counted_at 2026-10-17T11:56:11.486Z — spec_sources.verified_at for active_params_b — when this figure was counted; a refresh that keeps it does not rewrite it confidence 0.95 — spec_sources.confidence — from the tensor headers The parameters read to produce one token. It sets the decode speed; on a dense model it equals the total. Shown to one decimal: the underlying tensor count is exact, but more digits would claim a precision the published figure does not have (SPEC §14). active	UD-Q3_K_XL 0.97 GB (observed) OBSERVEDUD-Q3_K_XL file size on diskMeasured or published, passed through unchanged — not computed here. quantization file size — published at the source, passed through unchanged; nothing is computed here 0.97 GB, as recorded from https://huggingface.co/unsloth/Qwen3-VL-2B-Instruct-1M-GGUF/tree/main, file Qwen3-VL-2B-Instruct-1M-UD-Q3_K_XL.gguf quantization_file_size 0.97 GB — https://huggingface.co/unsloth/Qwen3-VL-2B-Instruct-1M-GGUF/tree/main, file Qwen3-VL-2B-Instruct-1M-UD-Q3_K_XL.gguf verified_at 2026-10-17T11:56:11.486Z — the row's own verification timestamp quantization UD-Q3_K_XL — gguf bits_per_weight 4.505 bits — the quantizer's own naming, or measured from the file The on-disk size of the weights. It is not the memory the model occupies: add the KV cache and the runtime overhead (SPEC §8.1). Preferred over the bits-per-weight formula wherever it exists, because it includes embeddings and any higher-precision tensors. Observed, not derived: this app records the figure and its source, and does not re-measure it. on disk	4.80 GB (estimate) ESTIMATEQwen3-VL 2B Instruct memory requirement at 32K contextArithmetic that also depends on a tunable assumption, listed below. required_gb = weights_gb + kv_cache_gb + overhead_gb required_gb = 0.97 + 3.76 + 0.08 = 4.8 GB weights_gb 0.968900448 GB — real file size kv_cache_gb 3.758096384 GB — kv formula overhead_gb 0.07751203584 GB — overhead fraction Real file size used instead of the params × bits formula (SPEC §8.1). 28 layers keep a full KV cache. Runtime buffers, activations, and allocator slack approximated as a fraction of weights.	89 of 89 (derived) DERIVEDmachines that run Qwen3-VL 2B Instruct at ≥ 15 tok/sArithmetic on observed inputs only. No tunable assumption enters it. machines = #{candidate rigs where the model loads AND clears the Plan's speed floor} 89 evaluated → 89 hold it in memory → 89 of those also reach ≥ 15 tok/s candidates_evaluated 89 machines — every complete rig the search could build from the priced catalog hold_it_in_memory 89 machines — required memory ≤ usable memory (SPEC §8.6) clear_the_speed_floor 89 machines — ≥ 15 tok/s The denominator is the machines this catalog can build today, not every machine that exists. A model with no shape or no sized quantization is counted on no machine at all, and says so rather than counting as zero.	SIZEABLE	read stale 3 d — past the 36 hour freshness window — Qwen3-VL 2B Instruct: the least recent read of its data from source (shape counted Oct 17, 7:56 AM ET from the tensor headers of https://huggingface.co/Qwen/Qwen3-VL-2B-Instruct; quantization sizes read Oct 17, 7:56 AM ET); never checked upstream.
Qwen3-ASR 0.6BAlibaba Qwen · dense · speech to text	0.8 (derived) DERIVEDQwen3-ASR 0.6B total parametersArithmetic on observed inputs only. No tunable assumption enters it. Σ params of language-model tensors — Cross-checked: tensor count 0.94B vs Hugging Face 0.94B (0%). 0.8 B, counted from the tensor headers of https://huggingface.co/Qwen/Qwen3-ASR-0.6B at 2026-10-17T11:50:54.554Z total_parameters 0.8 B — https://huggingface.co/Qwen/Qwen3-ASR-0.6B counted_at 2026-10-17T11:50:54.554Z — spec_sources.verified_at for total_params_b — when this figure was counted; a refresh that keeps it does not rewrite it confidence 0.95 — spec_sources.confidence — from the tensor headers The parameters of the language-model tensors only, counted from the tensor headers: a vision or audio tower and multi-token-prediction heads are excluded, so it can be below the whole-checkpoint count the formula's cross-check quotes. It sizes the weights only when no measured file size is stored for the chosen quantization; it is not what is read per token on a mixture-of-experts model. Shown to one decimal: the underlying tensor count is exact, but more digits would claim a precision the published figure does not have (SPEC §14). 0.6 (derived) DERIVEDQwen3-ASR 0.6B active parameters per tokenArithmetic on observed inputs only. No tunable assumption enters it. active − input embedding table — Per-token memory read used for decode bandwidth (SPEC §8.2). 0.6 B, counted from the tensor headers of https://huggingface.co/Qwen/Qwen3-ASR-0.6B at 2026-10-17T11:50:54.554Z active_parameters_per_token 0.6 B — https://huggingface.co/Qwen/Qwen3-ASR-0.6B counted_at 2026-10-17T11:50:54.554Z — spec_sources.verified_at for active_params_b — when this figure was counted; a refresh that keeps it does not rewrite it confidence 0.95 — spec_sources.confidence — from the tensor headers The parameters read to produce one token. It sets the decode speed; on a dense model it equals the total. Shown to one decimal: the underlying tensor count is exact, but more digits would claim a precision the published figure does not have (SPEC §14). active	Q8_0 0.80 GB (observed) OBSERVEDQ8_0 file size on diskMeasured or published, passed through unchanged — not computed here. quantization file size — published at the source, passed through unchanged; nothing is computed here 0.80 GB, as recorded from https://huggingface.co/ggml-org/Qwen3-ASR-0.6B-GGUF/tree/main, file Qwen3-ASR-0.6B-Q8_0.gguf quantization_file_size 0.80 GB — https://huggingface.co/ggml-org/Qwen3-ASR-0.6B-GGUF/tree/main, file Qwen3-ASR-0.6B-Q8_0.gguf verified_at 2026-10-17T11:50:54.554Z — the row's own verification timestamp quantization Q8_0 — gguf bits_per_weight 8.5653 bits — the quantizer's own naming, or measured from the file The on-disk size of the weights. It is not the memory the model occupies: add the KV cache and the runtime overhead (SPEC §8.1). Preferred over the bits-per-weight formula wherever it exists, because it includes embeddings and any higher-precision tensors. Observed, not derived: this app records the figure and its source, and does not re-measure it. on disk	4.63 GB (estimate) ESTIMATEQwen3-ASR 0.6B memory requirement at 32K contextArithmetic that also depends on a tunable assumption, listed below. required_gb = weights_gb + kv_cache_gb + overhead_gb required_gb = 0.8 + 3.76 + 0.06 = 4.63 GB weights_gb 0.804749248 GB — real file size kv_cache_gb 3.758096384 GB — kv formula overhead_gb 0.06437993984 GB — overhead fraction Real file size used instead of the params × bits formula (SPEC §8.1). 28 layers keep a full KV cache. Runtime buffers, activations, and allocator slack approximated as a fraction of weights.	89 of 89 (derived) DERIVEDmachines that run Qwen3-ASR 0.6B at ≥ 15 tok/sArithmetic on observed inputs only. No tunable assumption enters it. machines = #{candidate rigs where the model loads AND clears the Plan's speed floor} 89 evaluated → 89 hold it in memory → 89 of those also reach ≥ 15 tok/s candidates_evaluated 89 machines — every complete rig the search could build from the priced catalog hold_it_in_memory 89 machines — required memory ≤ usable memory (SPEC §8.6) clear_the_speed_floor 89 machines — ≥ 15 tok/s The denominator is the machines this catalog can build today, not every machine that exists. A model with no shape or no sized quantization is counted on no machine at all, and says so rather than counting as zero.	SIZEABLE	read stale 3 d — past the 36 hour freshness window — Qwen3-ASR 0.6B: the least recent read of its data from source (shape counted Oct 17, 7:50 AM ET from the tensor headers of https://huggingface.co/Qwen/Qwen3-ASR-0.6B; quantization sizes read Oct 17, 7:50 AM ET); never checked upstream.
MiniCPM-V 4.6OpenBMB · dense · vision · 256K ctx max	0.8 (derived) DERIVEDMiniCPM-V 4.6 total parametersArithmetic on observed inputs only. No tunable assumption enters it. Σ params of language-model tensors — Cross-checked: tensor count 1.3B vs Hugging Face 1.3B (0%). 0.8 B, counted from the tensor headers of https://huggingface.co/openbmb/MiniCPM-V-4.6 at 2026-10-17T12:41:14.877Z total_parameters 0.8 B — https://huggingface.co/openbmb/MiniCPM-V-4.6 counted_at 2026-10-17T12:41:14.877Z — spec_sources.verified_at for total_params_b — when this figure was counted; a refresh that keeps it does not rewrite it confidence 0.95 — spec_sources.confidence — from the tensor headers The parameters of the language-model tensors only, counted from the tensor headers: a vision or audio tower and multi-token-prediction heads are excluded, so it can be below the whole-checkpoint count the formula's cross-check quotes. It sizes the weights only when no measured file size is stored for the chosen quantization; it is not what is read per token on a mixture-of-experts model. Shown to one decimal: the underlying tensor count is exact, but more digits would claim a precision the published figure does not have (SPEC §14). 0.5 (derived) DERIVEDMiniCPM-V 4.6 active parameters per tokenArithmetic on observed inputs only. No tunable assumption enters it. active − input embedding table — Per-token memory read used for decode bandwidth (SPEC §8.2). 0.5 B, counted from the tensor headers of https://huggingface.co/openbmb/MiniCPM-V-4.6 at 2026-10-17T12:41:14.877Z active_parameters_per_token 0.5 B — https://huggingface.co/openbmb/MiniCPM-V-4.6 counted_at 2026-10-17T12:41:14.877Z — spec_sources.verified_at for active_params_b — when this figure was counted; a refresh that keeps it does not rewrite it confidence 0.95 — spec_sources.confidence — from the tensor headers The parameters read to produce one token. It sets the decode speed; on a dense model it equals the total. Shown to one decimal: the underlying tensor count is exact, but more digits would claim a precision the published figure does not have (SPEC §14). active	Q4_K_M 0.53 GB (observed) OBSERVEDQ4_K_M file size on diskMeasured or published, passed through unchanged — not computed here. quantization file size — published at the source, passed through unchanged; nothing is computed here 0.53 GB, as recorded from https://huggingface.co/ggml-org/MiniCPM-V-4.6-GGUF/tree/main, file MiniCPM-V-4.6-Q4_K_M.gguf quantization_file_size 0.53 GB — https://huggingface.co/ggml-org/MiniCPM-V-4.6-GGUF/tree/main, file MiniCPM-V-4.6-Q4_K_M.gguf verified_at 2026-10-17T12:41:14.877Z — the row's own verification timestamp quantization Q4_K_M — gguf bits_per_weight 5.4397 bits — the quantizer's own naming, or measured from the file The on-disk size of the weights. It is not the memory the model occupies: add the KV cache and the runtime overhead (SPEC §8.1). Preferred over the bits-per-weight formula wherever it exists, because it includes embeddings and any higher-precision tensors. Observed, not derived: this app records the figure and its source, and does not re-measure it. on disk	0.97 GB (estimate) ESTIMATEMiniCPM-V 4.6 memory requirement at 32K contextArithmetic that also depends on a tunable assumption, listed below. required_gb = weights_gb + kv_cache_gb + overhead_gb required_gb = 0.53 + 0.4 + 0.04 = 0.97 GB weights_gb 0.529101536 GB — real file size kv_cache_gb 0.402653184 GB — kv formula overhead_gb 0.04232812288 GB — overhead fraction Real file size used instead of the params × bits formula (SPEC §8.1). 6 layers keep a full KV cache, 18 linear-attention/SSM layers keep only a constant-size state (treated as negligible). Runtime buffers, activations, and allocator slack approximated as a fraction of weights.	89 of 89 (derived) DERIVEDmachines that run MiniCPM-V 4.6 at ≥ 15 tok/sArithmetic on observed inputs only. No tunable assumption enters it. machines = #{candidate rigs where the model loads AND clears the Plan's speed floor} 89 evaluated → 89 hold it in memory → 89 of those also reach ≥ 15 tok/s candidates_evaluated 89 machines — every complete rig the search could build from the priced catalog hold_it_in_memory 89 machines — required memory ≤ usable memory (SPEC §8.6) clear_the_speed_floor 89 machines — ≥ 15 tok/s The denominator is the machines this catalog can build today, not every machine that exists. A model with no shape or no sized quantization is counted on no machine at all, and says so rather than counting as zero.	SIZEABLE	read stale 3 d — past the 36 hour freshness window — MiniCPM-V 4.6: the least recent read of its data from source (shape counted Oct 17, 8:41 AM ET from the tensor headers of https://huggingface.co/openbmb/MiniCPM-V-4.6; quantization sizes read Oct 17, 8:41 AM ET); never checked upstream.
gpt-oss-20bOpenAI · MoE · reasoning · 128K ctx max	20.9 (derived) DERIVEDgpt-oss-20b total parametersArithmetic on observed inputs only. No tunable assumption enters it. Σ params of language-model tensors — Cross-checked: tensor count 20.91B vs Hugging Face 20.91B (0%). 20.9 B, counted from the tensor headers of https://huggingface.co/openai/gpt-oss-20b at 2026-10-17T10:37:50.301Z total_parameters 20.9 B — https://huggingface.co/openai/gpt-oss-20b counted_at 2026-10-17T10:37:50.301Z — spec_sources.verified_at for total_params_b — when this figure was counted; a refresh that keeps it does not rewrite it confidence 0.95 — spec_sources.confidence — from the tensor headers The parameters of the language-model tensors only, counted from the tensor headers: a vision or audio tower and multi-token-prediction heads are excluded, so it can be below the whole-checkpoint count the formula's cross-check quotes. It sizes the weights only when no measured file size is stored for the chosen quantization; it is not what is read per token on a mixture-of-experts model. Shown to one decimal: the underlying tensor count is exact, but more digits would claim a precision the published figure does not have (SPEC §14). 3.6 (derived) DERIVEDgpt-oss-20b active parameters per tokenArithmetic on observed inputs only. No tunable assumption enters it. active − input embedding table — Per-token memory read used for decode bandwidth (SPEC §8.2). 3.6 B, counted from the tensor headers of https://huggingface.co/openai/gpt-oss-20b at 2026-10-17T10:37:50.301Z active_parameters_per_token 3.6 B — https://huggingface.co/openai/gpt-oss-20b counted_at 2026-10-17T10:37:50.301Z — spec_sources.verified_at for active_params_b — when this figure was counted; a refresh that keeps it does not rewrite it confidence 0.95 — spec_sources.confidence — from the tensor headers The parameters read to produce one token. It sets the decode speed; on a dense model it equals the total. Shown to one decimal: the underlying tensor count is exact, but more digits would claim a precision the published figure does not have (SPEC §14). active	Q2_K_L 11.8 GB (observed) OBSERVEDQ2_K_L file size on diskMeasured or published, passed through unchanged — not computed here. quantization file size — published at the source, passed through unchanged; nothing is computed here 11.76 GB, as recorded from https://huggingface.co/unsloth/gpt-oss-20b-GGUF/tree/main, file gpt-oss-20b-Q2_K_L.gguf quantization_file_size 11.76 GB — https://huggingface.co/unsloth/gpt-oss-20b-GGUF/tree/main, file gpt-oss-20b-Q2_K_L.gguf verified_at 2026-10-17T10:37:50.301Z — the row's own verification timestamp quantization Q2_K_L — gguf bits_per_weight 4.4975 bits — the quantizer's own naming, or measured from the file The on-disk size of the weights. It is not the memory the model occupies: add the KV cache and the runtime overhead (SPEC §8.1). Preferred over the bits-per-weight formula wherever it exists, because it includes embeddings and any higher-precision tensors. Observed, not derived: this app records the figure and its source, and does not re-measure it. on disk	13.5 GB (estimate) ESTIMATEgpt-oss-20b memory requirement at 32K contextArithmetic that also depends on a tunable assumption, listed below. required_gb = weights_gb + kv_cache_gb + overhead_gb required_gb = 11.76 + 0.81 + 0.94 = 13.51 GB weights_gb 11.757884608 GB — real file size kv_cache_gb 0.808452096 GB — kv formula overhead_gb 0.94063076864 GB — overhead fraction Real file size used instead of the params × bits formula (SPEC §8.1). 12 layers keep a full KV cache, 12 sliding-window layers are capped at 128 tokens. Runtime buffers, activations, and allocator slack approximated as a fraction of weights.	88 of 89 (derived) DERIVEDmachines that run gpt-oss-20b at ≥ 15 tok/sArithmetic on observed inputs only. No tunable assumption enters it. machines = #{candidate rigs where the model loads AND clears the Plan's speed floor} 89 evaluated → 88 hold it in memory → 88 of those also reach ≥ 15 tok/s candidates_evaluated 89 machines — every complete rig the search could build from the priced catalog hold_it_in_memory 88 machines — required memory ≤ usable memory (SPEC §8.6) clear_the_speed_floor 88 machines — ≥ 15 tok/s The denominator is the machines this catalog can build today, not every machine that exists. A model with no shape or no sized quantization is counted on no machine at all, and says so rather than counting as zero.	SIZEABLE	read stale 3 d — past the 36 hour freshness window — gpt-oss-20b: the least recent read of its data from source (shape counted Oct 17, 6:37 AM ET from the tensor headers of https://huggingface.co/openai/gpt-oss-20b; quantization sizes read Oct 17, 6:37 AM ET); never checked upstream.
Kimi-VL A3B Thinking 2506Moonshot AI · MoE · vision · 128K ctx max	16.0 (derived) DERIVEDKimi-VL A3B Thinking 2506 total parametersArithmetic on observed inputs only. No tunable assumption enters it. Σ params of language-model tensors — Cross-checked: tensor count 16.41B vs Hugging Face 16.41B (0%). 16.0 B, counted from the tensor headers of https://huggingface.co/moonshotai/Kimi-VL-A3B-Thinking-2506 at 2026-10-17T12:54:45.786Z total_parameters 16.0 B — https://huggingface.co/moonshotai/Kimi-VL-A3B-Thinking-2506 counted_at 2026-10-17T12:54:45.786Z — spec_sources.verified_at for total_params_b — when this figure was counted; a refresh that keeps it does not rewrite it confidence 0.95 — spec_sources.confidence — from the tensor headers The parameters of the language-model tensors only, counted from the tensor headers: a vision or audio tower and multi-token-prediction heads are excluded, so it can be below the whole-checkpoint count the formula's cross-check quotes. It sizes the weights only when no measured file size is stored for the chosen quantization; it is not what is read per token on a mixture-of-experts model. Shown to one decimal: the underlying tensor count is exact, but more digits would claim a precision the published figure does not have (SPEC §14). 2.6 (derived) DERIVEDKimi-VL A3B Thinking 2506 active parameters per tokenArithmetic on observed inputs only. No tunable assumption enters it. active − input embedding table — Per-token memory read used for decode bandwidth (SPEC §8.2). 2.6 B, counted from the tensor headers of https://huggingface.co/moonshotai/Kimi-VL-A3B-Thinking-2506 at 2026-10-17T12:54:45.786Z active_parameters_per_token 2.6 B — https://huggingface.co/moonshotai/Kimi-VL-A3B-Thinking-2506 counted_at 2026-10-17T12:54:45.786Z — spec_sources.verified_at for active_params_b — when this figure was counted; a refresh that keeps it does not rewrite it confidence 0.95 — spec_sources.confidence — from the tensor headers The parameters read to produce one token. It sets the decode speed; on a dense model it equals the total. Shown to one decimal: the underlying tensor count is exact, but more digits would claim a precision the published figure does not have (SPEC §14). active	Q4_K_M 10.5 GB (observed) OBSERVEDQ4_K_M file size on diskMeasured or published, passed through unchanged — not computed here. quantization file size — published at the source, passed through unchanged; nothing is computed here 10.54 GB, as recorded from https://huggingface.co/ggml-org/Kimi-VL-A3B-Thinking-2506-GGUF/tree/main, file Kimi-VL-A3B-Thinking-2506-Q4_K_M.gguf quantization_file_size 10.54 GB — https://huggingface.co/ggml-org/Kimi-VL-A3B-Thinking-2506-GGUF/tree/main, file Kimi-VL-A3B-Thinking-2506-Q4_K_M.gguf verified_at 2026-10-17T12:54:45.786Z — the row's own verification timestamp quantization Q4_K_M — gguf bits_per_weight 5.2835 bits — the quantizer's own naming, or measured from the file The on-disk size of the weights. It is not the memory the model occupies: add the KV cache and the runtime overhead (SPEC §8.1). Preferred over the bits-per-weight formula wherever it exists, because it includes embeddings and any higher-precision tensors. Observed, not derived: this app records the figure and its source, and does not re-measure it. on disk	12.4 GB (estimate) ESTIMATEKimi-VL A3B Thinking 2506 memory requirement at 32K contextArithmetic that also depends on a tunable assumption, listed below. required_gb = weights_gb + kv_cache_gb + overhead_gb required_gb = 10.54 + 1.02 + 0.84 = 12.4 GB weights_gb 10.54074768 GB — real file size kv_cache_gb 1.019215872 GB — kv formula overhead_gb 0.8432598144000001 GB — overhead fraction Real file size used instead of the params × bits formula (SPEC §8.1). 27 layers keep a full KV cache. Runtime buffers, activations, and allocator slack approximated as a fraction of weights.	88 of 89 (derived) DERIVEDmachines that run Kimi-VL A3B Thinking 2506 at ≥ 15 tok/sArithmetic on observed inputs only. No tunable assumption enters it. machines = #{candidate rigs where the model loads AND clears the Plan's speed floor} 89 evaluated → 88 hold it in memory → 88 of those also reach ≥ 15 tok/s candidates_evaluated 89 machines — every complete rig the search could build from the priced catalog hold_it_in_memory 88 machines — required memory ≤ usable memory (SPEC §8.6) clear_the_speed_floor 88 machines — ≥ 15 tok/s The denominator is the machines this catalog can build today, not every machine that exists. A model with no shape or no sized quantization is counted on no machine at all, and says so rather than counting as zero.	SIZEABLE	read stale 3 d — past the 36 hour freshness window — Kimi-VL A3B Thinking 2506: the least recent read of its data from source (shape counted Oct 17, 8:54 AM ET from the tensor headers of https://huggingface.co/moonshotai/Kimi-VL-A3B-Thinking-2506; quantization sizes read Oct 17, 8:54 AM ET); never checked upstream.
Qwen3.6-35B-A3BAlibaba Qwen · MoE · coding, vision · 256K ctx max	34.7 (derived) DERIVEDQwen3.6-35B-A3B total parametersArithmetic on observed inputs only. No tunable assumption enters it. Σ params of language-model tensors — Cross-checked: tensor count 35.95B vs Hugging Face 35.95B (0%). 34.7 B, counted from the tensor headers of https://huggingface.co/Qwen/Qwen3.6-35B-A3B at 2026-10-18T07:39:13.960Z total_parameters 34.7 B — https://huggingface.co/Qwen/Qwen3.6-35B-A3B counted_at 2026-10-18T07:39:13.960Z — spec_sources.verified_at for total_params_b — when this figure was counted; a refresh that keeps it does not rewrite it confidence 0.95 — spec_sources.confidence — from the tensor headers The parameters of the language-model tensors only, counted from the tensor headers: a vision or audio tower and multi-token-prediction heads are excluded, so it can be below the whole-checkpoint count the formula's cross-check quotes. It sizes the weights only when no measured file size is stored for the chosen quantization; it is not what is read per token on a mixture-of-experts model. Shown to one decimal: the underlying tensor count is exact, but more digits would claim a precision the published figure does not have (SPEC §14). 2.9 (derived) DERIVEDQwen3.6-35B-A3B active parameters per tokenArithmetic on observed inputs only. No tunable assumption enters it. active − input embedding table — Per-token memory read used for decode bandwidth (SPEC §8.2). 2.9 B, counted from the tensor headers of https://huggingface.co/Qwen/Qwen3.6-35B-A3B at 2026-10-18T07:39:13.960Z active_parameters_per_token 2.9 B — https://huggingface.co/Qwen/Qwen3.6-35B-A3B counted_at 2026-10-18T07:39:13.960Z — spec_sources.verified_at for active_params_b — when this figure was counted; a refresh that keeps it does not rewrite it confidence 0.95 — spec_sources.confidence — from the tensor headers The parameters read to produce one token. It sets the decode speed; on a dense model it equals the total. Shown to one decimal: the underlying tensor count is exact, but more digits would claim a precision the published figure does not have (SPEC §14). active	IQ4_XS 19.7 GB (observed) OBSERVEDIQ4_XS file size on diskMeasured or published, passed through unchanged — not computed here. quantization file size — published at the source, passed through unchanged; nothing is computed here 19.70 GB, as recorded from https://huggingface.co/bartowski/Qwen_Qwen3.6-35B-A3B-GGUF/tree/main, file Qwen_Qwen3.6-35B-A3B-IQ4_XS.gguf quantization_file_size 19.70 GB — https://huggingface.co/bartowski/Qwen_Qwen3.6-35B-A3B-GGUF/tree/main, file Qwen_Qwen3.6-35B-A3B-IQ4_XS.gguf verified_at 2026-10-18T07:39:13.960Z — the row's own verification timestamp quantization IQ4_XS — gguf bits_per_weight 4.5468 bits — the quantizer's own naming, or measured from the file The on-disk size of the weights. It is not the memory the model occupies: add the KV cache and the runtime overhead (SPEC §8.1). Preferred over the bits-per-weight formula wherever it exists, because it includes embeddings and any higher-precision tensors. Observed, not derived: this app records the figure and its source, and does not re-measure it. on disk	21.9 GB (estimate) ESTIMATEQwen3.6-35B-A3B memory requirement at 32K contextArithmetic that also depends on a tunable assumption, listed below. required_gb = weights_gb + kv_cache_gb + overhead_gb required_gb = 19.7 + 0.67 + 1.58 = 21.95 GB weights_gb 19.69955392 GB — real file size kv_cache_gb 0.67108864 GB — kv formula overhead_gb 1.5759643136 GB — overhead fraction Real file size used instead of the params × bits formula (SPEC §8.1). 10 layers keep a full KV cache, 30 linear-attention/SSM layers keep only a constant-size state (treated as negligible). Runtime buffers, activations, and allocator slack approximated as a fraction of weights.	86 of 89 (derived) DERIVEDmachines that run Qwen3.6-35B-A3B at ≥ 15 tok/sArithmetic on observed inputs only. No tunable assumption enters it. machines = #{candidate rigs where the model loads AND clears the Plan's speed floor} 89 evaluated → 86 hold it in memory → 86 of those also reach ≥ 15 tok/s candidates_evaluated 89 machines — every complete rig the search could build from the priced catalog hold_it_in_memory 86 machines — required memory ≤ usable memory (SPEC §8.6) clear_the_speed_floor 86 machines — ≥ 15 tok/s The denominator is the machines this catalog can build today, not every machine that exists. A model with no shape or no sized quantization is counted on no machine at all, and says so rather than counting as zero.	SIZEABLE	read stale 2 d — past the 36 hour freshness window — Qwen3.6-35B-A3B: the least recent read of its data from source (shape counted Oct 18, 3:39 AM ET from the tensor headers of https://huggingface.co/Qwen/Qwen3.6-35B-A3B; quantization sizes read Oct 18, 3:39 AM ET); no upstream check after it.
Gemma 4 12B Unified ITGoogle · dense · vision · 256K ctx max	11.9 (derived) DERIVEDGemma 4 12B Unified IT total parametersArithmetic on observed inputs only. No tunable assumption enters it. Σ params of language-model tensors — Cross-checked: tensor count 11.96B vs Hugging Face 11.96B (0%). 11.9 B, counted from the tensor headers of https://huggingface.co/google/gemma-4-12B-it at 2026-10-17T12:59:28.666Z total_parameters 11.9 B — https://huggingface.co/google/gemma-4-12B-it counted_at 2026-10-17T12:59:28.666Z — spec_sources.verified_at for total_params_b — when this figure was counted; a refresh that keeps it does not rewrite it confidence 0.95 — spec_sources.confidence — from the tensor headers The parameters of the language-model tensors only, counted from the tensor headers: a vision or audio tower and multi-token-prediction heads are excluded, so it can be below the whole-checkpoint count the formula's cross-check quotes. It sizes the weights only when no measured file size is stored for the chosen quantization; it is not what is read per token on a mixture-of-experts model. Shown to one decimal: the underlying tensor count is exact, but more digits would claim a precision the published figure does not have (SPEC §14). 10.9 (derived) DERIVEDGemma 4 12B Unified IT active parameters per tokenArithmetic on observed inputs only. No tunable assumption enters it. active − input embedding table — Per-token memory read used for decode bandwidth (SPEC §8.2). 10.9 B, counted from the tensor headers of https://huggingface.co/google/gemma-4-12B-it at 2026-10-17T12:59:28.666Z active_parameters_per_token 10.9 B — https://huggingface.co/google/gemma-4-12B-it counted_at 2026-10-17T12:59:28.666Z — spec_sources.verified_at for active_params_b — when this figure was counted; a refresh that keeps it does not rewrite it confidence 0.95 — spec_sources.confidence — from the tensor headers The parameters read to produce one token. It sets the decode speed; on a dense model it equals the total. Shown to one decimal: the underlying tensor count is exact, but more digits would claim a precision the published figure does not have (SPEC §14). active	IQ4_NL 6.72 GB (observed) OBSERVEDIQ4_NL file size on diskMeasured or published, passed through unchanged — not computed here. quantization file size — published at the source, passed through unchanged; nothing is computed here 6.72 GB, as recorded from https://huggingface.co/unsloth/gemma-4-12b-it-GGUF/tree/main, file gemma-4-12b-it-IQ4_NL.gguf quantization_file_size 6.72 GB — https://huggingface.co/unsloth/gemma-4-12b-it-GGUF/tree/main, file gemma-4-12b-it-IQ4_NL.gguf verified_at 2026-10-17T12:59:28.666Z — the row's own verification timestamp quantization IQ4_NL — gguf bits_per_weight 4.4991 bits — the quantizer's own naming, or measured from the file The on-disk size of the weights. It is not the memory the model occupies: add the KV cache and the runtime overhead (SPEC §8.1). Preferred over the bits-per-weight formula wherever it exists, because it includes embeddings and any higher-precision tensors. Observed, not derived: this app records the figure and its source, and does not re-measure it. on disk	7.86 GB (estimate) ESTIMATEGemma 4 12B Unified IT memory requirement at 32K contextArithmetic that also depends on a tunable assumption, listed below. required_gb = weights_gb + kv_cache_gb + overhead_gb required_gb = 6.72 + 0.6 + 0.54 = 7.86 GB weights_gb 6.71635744 GB — real file size kv_cache_gb 0.603979776 GB — kv formula overhead_gb 0.5373085952000001 GB — overhead fraction Real file size used instead of the params × bits formula (SPEC §8.1). 8 layers keep a full KV cache, 40 sliding-window layers are capped at 1024 tokens. Runtime buffers, activations, and allocator slack approximated as a fraction of weights.	85 of 89 (derived) DERIVEDmachines that run Gemma 4 12B Unified IT at ≥ 15 tok/sArithmetic on observed inputs only. No tunable assumption enters it. machines = #{candidate rigs where the model loads AND clears the Plan's speed floor} 89 evaluated → 89 hold it in memory → 85 of those also reach ≥ 15 tok/s candidates_evaluated 89 machines — every complete rig the search could build from the priced catalog hold_it_in_memory 89 machines — required memory ≤ usable memory (SPEC §8.6) clear_the_speed_floor 85 machines — ≥ 15 tok/s The denominator is the machines this catalog can build today, not every machine that exists. A model with no shape or no sized quantization is counted on no machine at all, and says so rather than counting as zero.	SIZEABLE	read stale 3 d — past the 36 hour freshness window — Gemma 4 12B Unified IT: the least recent read of its data from source (shape counted Oct 17, 8:59 AM ET from the tensor headers of https://huggingface.co/google/gemma-4-12B-it; quantization sizes read Oct 17, 8:59 AM ET); never checked upstream.
Qwen3-VL 4B InstructAlibaba Qwen · dense · vision · 256K ctx max	4.0 (derived) DERIVEDQwen3-VL 4B Instruct total parametersArithmetic on observed inputs only. No tunable assumption enters it. Σ params of language-model tensors — Cross-checked: tensor count 4.44B vs Hugging Face 4.44B (0%). 4.0 B, counted from the tensor headers of https://huggingface.co/Qwen/Qwen3-VL-4B-Instruct at 2026-10-17T11:57:33.713Z total_parameters 4.0 B — https://huggingface.co/Qwen/Qwen3-VL-4B-Instruct counted_at 2026-10-17T11:57:33.713Z — spec_sources.verified_at for total_params_b — when this figure was counted; a refresh that keeps it does not rewrite it confidence 0.95 — spec_sources.confidence — from the tensor headers The parameters of the language-model tensors only, counted from the tensor headers: a vision or audio tower and multi-token-prediction heads are excluded, so it can be below the whole-checkpoint count the formula's cross-check quotes. It sizes the weights only when no measured file size is stored for the chosen quantization; it is not what is read per token on a mixture-of-experts model. Shown to one decimal: the underlying tensor count is exact, but more digits would claim a precision the published figure does not have (SPEC §14). 3.6 (derived) DERIVEDQwen3-VL 4B Instruct active parameters per tokenArithmetic on observed inputs only. No tunable assumption enters it. active − input embedding table — Per-token memory read used for decode bandwidth (SPEC §8.2). 3.6 B, counted from the tensor headers of https://huggingface.co/Qwen/Qwen3-VL-4B-Instruct at 2026-10-17T11:57:33.713Z active_parameters_per_token 3.6 B — https://huggingface.co/Qwen/Qwen3-VL-4B-Instruct counted_at 2026-10-17T11:57:33.713Z — spec_sources.verified_at for active_params_b — when this figure was counted; a refresh that keeps it does not rewrite it confidence 0.95 — spec_sources.confidence — from the tensor headers The parameters read to produce one token. It sets the decode speed; on a dense model it equals the total. Shown to one decimal: the underlying tensor count is exact, but more digits would claim a precision the published figure does not have (SPEC §14). active	IQ4_XS 2.27 GB (observed) OBSERVEDIQ4_XS file size on diskMeasured or published, passed through unchanged — not computed here. quantization file size — published at the source, passed through unchanged; nothing is computed here 2.27 GB, as recorded from https://huggingface.co/bartowski/Qwen_Qwen3-VL-4B-Instruct-GGUF/tree/main, file Qwen_Qwen3-VL-4B-Instruct-IQ4_XS.gguf quantization_file_size 2.27 GB — https://huggingface.co/bartowski/Qwen_Qwen3-VL-4B-Instruct-GGUF/tree/main, file Qwen_Qwen3-VL-4B-Instruct-IQ4_XS.gguf verified_at 2026-10-17T11:57:33.713Z — the row's own verification timestamp quantization IQ4_XS — gguf bits_per_weight 4.5161 bits — the quantizer's own naming, or measured from the file The on-disk size of the weights. It is not the memory the model occupies: add the KV cache and the runtime overhead (SPEC §8.1). Preferred over the bits-per-weight formula wherever it exists, because it includes embeddings and any higher-precision tensors. Observed, not derived: this app records the figure and its source, and does not re-measure it. on disk	7.28 GB (estimate) ESTIMATEQwen3-VL 4B Instruct memory requirement at 32K contextArithmetic that also depends on a tunable assumption, listed below. required_gb = weights_gb + kv_cache_gb + overhead_gb required_gb = 2.27 + 4.83 + 0.18 = 7.28 GB weights_gb 2.270752672 GB — real file size kv_cache_gb 4.831838208 GB — kv formula overhead_gb 0.18166021376 GB — overhead fraction Real file size used instead of the params × bits formula (SPEC §8.1). 36 layers keep a full KV cache. Runtime buffers, activations, and allocator slack approximated as a fraction of weights.	85 of 89 (derived) DERIVEDmachines that run Qwen3-VL 4B Instruct at ≥ 15 tok/sArithmetic on observed inputs only. No tunable assumption enters it. machines = #{candidate rigs where the model loads AND clears the Plan's speed floor} 89 evaluated → 89 hold it in memory → 85 of those also reach ≥ 15 tok/s candidates_evaluated 89 machines — every complete rig the search could build from the priced catalog hold_it_in_memory 89 machines — required memory ≤ usable memory (SPEC §8.6) clear_the_speed_floor 85 machines — ≥ 15 tok/s The denominator is the machines this catalog can build today, not every machine that exists. A model with no shape or no sized quantization is counted on no machine at all, and says so rather than counting as zero.	SIZEABLE	read stale 3 d — past the 36 hour freshness window — Qwen3-VL 4B Instruct: the least recent read of its data from source (shape counted Oct 17, 7:57 AM ET from the tensor headers of https://huggingface.co/Qwen/Qwen3-VL-4B-Instruct; quantization sizes read Oct 17, 7:57 AM ET); never checked upstream.
GLM-4.6V-FlashZ.ai · dense · vision · 128K ctx max	9.4 (derived) DERIVEDGLM-4.6V-Flash total parametersArithmetic on observed inputs only. No tunable assumption enters it. Σ params of language-model tensors — Cross-checked: tensor count 10.29B vs Hugging Face 10.29B (0%). 9.4 B, counted from the tensor headers of https://huggingface.co/zai-org/GLM-4.6V-Flash at 2026-10-17T12:52:43.216Z total_parameters 9.4 B — https://huggingface.co/zai-org/GLM-4.6V-Flash counted_at 2026-10-17T12:52:43.216Z — spec_sources.verified_at for total_params_b — when this figure was counted; a refresh that keeps it does not rewrite it confidence 0.95 — spec_sources.confidence — from the tensor headers The parameters of the language-model tensors only, counted from the tensor headers: a vision or audio tower and multi-token-prediction heads are excluded, so it can be below the whole-checkpoint count the formula's cross-check quotes. It sizes the weights only when no measured file size is stored for the chosen quantization; it is not what is read per token on a mixture-of-experts model. Shown to one decimal: the underlying tensor count is exact, but more digits would claim a precision the published figure does not have (SPEC §14). 8.8 (derived) DERIVEDGLM-4.6V-Flash active parameters per tokenArithmetic on observed inputs only. No tunable assumption enters it. active − input embedding table — Per-token memory read used for decode bandwidth (SPEC §8.2). 8.8 B, counted from the tensor headers of https://huggingface.co/zai-org/GLM-4.6V-Flash at 2026-10-17T12:52:43.216Z active_parameters_per_token 8.8 B — https://huggingface.co/zai-org/GLM-4.6V-Flash counted_at 2026-10-17T12:52:43.216Z — spec_sources.verified_at for active_params_b — when this figure was counted; a refresh that keeps it does not rewrite it confidence 0.95 — spec_sources.confidence — from the tensor headers The parameters read to produce one token. It sets the decode speed; on a dense model it equals the total. Shown to one decimal: the underlying tensor count is exact, but more digits would claim a precision the published figure does not have (SPEC §14). active	IQ4_XS 5.28 GB (observed) OBSERVEDIQ4_XS file size on diskMeasured or published, passed through unchanged — not computed here. quantization file size — published at the source, passed through unchanged; nothing is computed here 5.28 GB, as recorded from https://huggingface.co/unsloth/GLM-4.6V-Flash-GGUF/tree/main, file GLM-4.6V-Flash-IQ4_XS.gguf quantization_file_size 5.28 GB — https://huggingface.co/unsloth/GLM-4.6V-Flash-GGUF/tree/main, file GLM-4.6V-Flash-IQ4_XS.gguf verified_at 2026-10-17T12:52:43.216Z — the row's own verification timestamp quantization IQ4_XS — gguf bits_per_weight 4.4949 bits — the quantizer's own naming, or measured from the file The on-disk size of the weights. It is not the memory the model occupies: add the KV cache and the runtime overhead (SPEC §8.1). Preferred over the bits-per-weight formula wherever it exists, because it includes embeddings and any higher-precision tensors. Observed, not derived: this app records the figure and its source, and does not re-measure it. on disk	7.05 GB (estimate) ESTIMATEGLM-4.6V-Flash memory requirement at 32K contextArithmetic that also depends on a tunable assumption, listed below. required_gb = weights_gb + kv_cache_gb + overhead_gb required_gb = 5.28 + 1.34 + 0.42 = 7.05 GB weights_gb 5.281678944 GB — real file size kv_cache_gb 1.34217728 GB — kv formula overhead_gb 0.42253431552000004 GB — overhead fraction Real file size used instead of the params × bits formula (SPEC §8.1). 40 layers keep a full KV cache. Runtime buffers, activations, and allocator slack approximated as a fraction of weights.	85 of 89 (derived) DERIVEDmachines that run GLM-4.6V-Flash at ≥ 15 tok/sArithmetic on observed inputs only. No tunable assumption enters it. machines = #{candidate rigs where the model loads AND clears the Plan's speed floor} 89 evaluated → 89 hold it in memory → 85 of those also reach ≥ 15 tok/s candidates_evaluated 89 machines — every complete rig the search could build from the priced catalog hold_it_in_memory 89 machines — required memory ≤ usable memory (SPEC §8.6) clear_the_speed_floor 85 machines — ≥ 15 tok/s The denominator is the machines this catalog can build today, not every machine that exists. A model with no shape or no sized quantization is counted on no machine at all, and says so rather than counting as zero.	SIZEABLE	read stale 3 d — past the 36 hour freshness window — GLM-4.6V-Flash: the least recent read of its data from source (shape counted Oct 17, 8:52 AM ET from the tensor headers of https://huggingface.co/zai-org/GLM-4.6V-Flash; quantization sizes read Oct 17, 8:52 AM ET); never checked upstream.
TRELLIS.2 4BMicrosoft · pipeline · 3d generation	Unavailable — TRELLIS.2 4B is a pipeline of several component models, not one network, so it has no single total parameters figure. Its size is the memory its maintainers publish for the whole pipeline. (missing: not applicable to a pipeline)	not quantized as one file	24.0 GB (observed: the maintainers’ published minimum)	82 of 89 (derived) DERIVEDmachines that run TRELLIS.2 4BArithmetic on observed inputs only. No tunable assumption enters it. machines = #{candidate rigs with enough memory AND a backend this pipeline supports} 89 evaluated → 82 hold it in memory → 82 of those have a backend it can run on candidates_evaluated 89 machines — every complete rig the search could build from the priced catalog hold_it_in_memory 82 machines — required memory ≤ usable memory (SPEC §8.6) supported_backend 82 machines — the CUDA / ROCm / Metal support the maintainers publish The denominator is the machines this catalog can build today, not every machine that exists. A model with no shape or no sized quantization is counted on no machine at all, and says so rather than counting as zero. A pipeline model has no token rate, so no speed floor is applied to it — only memory and backend support.	SIZEABLE	Unavailable — TRELLIS.2 4B: A pipeline model is sized by the memory and backends its maintainers publish, recorded in the seed file; nothing about it is read from Hugging Face, so it has no read date and no upstream check. (missing: not applicable to a pipeline)
InternVL3.5 30B-A3BShanghai AI Lab (OpenGVLab) · MoE · vision · 40K ctx max	30.5 (derived) DERIVEDInternVL3.5 30B-A3B total parametersArithmetic on observed inputs only. No tunable assumption enters it. Σ params of language-model tensors — Cross-checked: tensor count 30.85B vs Hugging Face 30.85B (0%). 30.5 B, counted from the tensor headers of https://huggingface.co/OpenGVLab/InternVL3_5-30B-A3B at 2026-10-17T12:33:31.930Z total_parameters 30.5 B — https://huggingface.co/OpenGVLab/InternVL3_5-30B-A3B counted_at 2026-10-17T12:33:31.930Z — spec_sources.verified_at for total_params_b — when this figure was counted; a refresh that keeps it does not rewrite it confidence 0.95 — spec_sources.confidence — from the tensor headers The parameters of the language-model tensors only, counted from the tensor headers: a vision or audio tower and multi-token-prediction heads are excluded, so it can be below the whole-checkpoint count the formula's cross-check quotes. It sizes the weights only when no measured file size is stored for the chosen quantization; it is not what is read per token on a mixture-of-experts model. Shown to one decimal: the underlying tensor count is exact, but more digits would claim a precision the published figure does not have (SPEC §14). 3.1 (derived) DERIVEDInternVL3.5 30B-A3B active parameters per tokenArithmetic on observed inputs only. No tunable assumption enters it. active − input embedding table — Per-token memory read used for decode bandwidth (SPEC §8.2). 3.1 B, counted from the tensor headers of https://huggingface.co/OpenGVLab/InternVL3_5-30B-A3B at 2026-10-17T12:33:31.930Z active_parameters_per_token 3.1 B — https://huggingface.co/OpenGVLab/InternVL3_5-30B-A3B counted_at 2026-10-17T12:33:31.930Z — spec_sources.verified_at for active_params_b — when this figure was counted; a refresh that keeps it does not rewrite it confidence 0.95 — spec_sources.confidence — from the tensor headers The parameters read to produce one token. It sets the decode speed; on a dense model it equals the total. Shown to one decimal: the underlying tensor count is exact, but more digits would claim a precision the published figure does not have (SPEC §14). active	IQ4_NL 17.4 GB (observed) OBSERVEDIQ4_NL file size on diskMeasured or published, passed through unchanged — not computed here. quantization file size — published at the source, passed through unchanged; nothing is computed here 17.39 GB, as recorded from https://huggingface.co/bartowski/OpenGVLab_InternVL3_5-30B-A3B-GGUF/tree/main, file OpenGVLab_InternVL3_5-30B-A3B-IQ4_NL.gguf quantization_file_size 17.39 GB — https://huggingface.co/bartowski/OpenGVLab_InternVL3_5-30B-A3B-GGUF/tree/main, file OpenGVLab_InternVL3_5-30B-A3B-IQ4_NL.gguf verified_at 2026-10-17T12:33:31.930Z — the row's own verification timestamp quantization IQ4_NL — gguf bits_per_weight 4.5537 bits — the quantizer's own naming, or measured from the file The on-disk size of the weights. It is not the memory the model occupies: add the KV cache and the runtime overhead (SPEC §8.1). Preferred over the bits-per-weight formula wherever it exists, because it includes embeddings and any higher-precision tensors. Observed, not derived: this app records the figure and its source, and does not re-measure it. on disk	22.0 GB (estimate) ESTIMATEInternVL3.5 30B-A3B memory requirement at 32K contextArithmetic that also depends on a tunable assumption, listed below. required_gb = weights_gb + kv_cache_gb + overhead_gb required_gb = 17.39 + 3.22 + 1.39 = 22 GB weights_gb 17.386275744 GB — real file size kv_cache_gb 3.221225472 GB — kv formula overhead_gb 1.39090205952 GB — overhead fraction Real file size used instead of the params × bits formula (SPEC §8.1). 48 layers keep a full KV cache. Runtime buffers, activations, and allocator slack approximated as a fraction of weights.	82 of 89 (derived) DERIVEDmachines that run InternVL3.5 30B-A3B at ≥ 15 tok/sArithmetic on observed inputs only. No tunable assumption enters it. machines = #{candidate rigs where the model loads AND clears the Plan's speed floor} 89 evaluated → 86 hold it in memory → 82 of those also reach ≥ 15 tok/s candidates_evaluated 89 machines — every complete rig the search could build from the priced catalog hold_it_in_memory 86 machines — required memory ≤ usable memory (SPEC §8.6) clear_the_speed_floor 82 machines — ≥ 15 tok/s The denominator is the machines this catalog can build today, not every machine that exists. A model with no shape or no sized quantization is counted on no machine at all, and says so rather than counting as zero.	SIZEABLE	read stale 3 d — past the 36 hour freshness window — InternVL3.5 30B-A3B: the least recent read of its data from source (shape counted Oct 17, 8:33 AM ET from the tensor headers of https://huggingface.co/OpenGVLab/InternVL3_5-30B-A3B; quantization sizes read Oct 17, 8:33 AM ET); never checked upstream.
Qwen3-VL 30B-A3B InstructAlibaba Qwen · MoE · vision · 256K ctx max	30.5 (derived) DERIVEDQwen3-VL 30B-A3B Instruct total parametersArithmetic on observed inputs only. No tunable assumption enters it. Σ params of language-model tensors — Cross-checked: tensor count 31.07B vs Hugging Face 31.07B (0%). 30.5 B, counted from the tensor headers of https://huggingface.co/Qwen/Qwen3-VL-30B-A3B-Instruct at 2026-10-17T12:06:36.631Z total_parameters 30.5 B — https://huggingface.co/Qwen/Qwen3-VL-30B-A3B-Instruct counted_at 2026-10-17T12:06:36.631Z — spec_sources.verified_at for total_params_b — when this figure was counted; a refresh that keeps it does not rewrite it confidence 0.95 — spec_sources.confidence — from the tensor headers The parameters of the language-model tensors only, counted from the tensor headers: a vision or audio tower and multi-token-prediction heads are excluded, so it can be below the whole-checkpoint count the formula's cross-check quotes. It sizes the weights only when no measured file size is stored for the chosen quantization; it is not what is read per token on a mixture-of-experts model. Shown to one decimal: the underlying tensor count is exact, but more digits would claim a precision the published figure does not have (SPEC §14). 3.0 (derived) DERIVEDQwen3-VL 30B-A3B Instruct active parameters per tokenArithmetic on observed inputs only. No tunable assumption enters it. active − input embedding table — Per-token memory read used for decode bandwidth (SPEC §8.2). 3.0 B, counted from the tensor headers of https://huggingface.co/Qwen/Qwen3-VL-30B-A3B-Instruct at 2026-10-17T12:06:36.631Z active_parameters_per_token 3.0 B — https://huggingface.co/Qwen/Qwen3-VL-30B-A3B-Instruct counted_at 2026-10-17T12:06:36.631Z — spec_sources.verified_at for active_params_b — when this figure was counted; a refresh that keeps it does not rewrite it confidence 0.95 — spec_sources.confidence — from the tensor headers The parameters read to produce one token. It sets the decode speed; on a dense model it equals the total. Shown to one decimal: the underlying tensor count is exact, but more digits would claim a precision the published figure does not have (SPEC §14). active	IQ4_NL 17.3 GB (observed) OBSERVEDIQ4_NL file size on diskMeasured or published, passed through unchanged — not computed here. quantization file size — published at the source, passed through unchanged; nothing is computed here 17.31 GB, as recorded from https://huggingface.co/unsloth/Qwen3-VL-30B-A3B-Instruct-1M-GGUF/tree/main, file Qwen3-VL-30B-A3B-Instruct-1M-IQ4_NL.gguf quantization_file_size 17.31 GB — https://huggingface.co/unsloth/Qwen3-VL-30B-A3B-Instruct-1M-GGUF/tree/main, file Qwen3-VL-30B-A3B-Instruct-1M-IQ4_NL.gguf verified_at 2026-10-17T12:06:36.631Z — the row's own verification timestamp quantization IQ4_NL — gguf bits_per_weight 4.5358 bits — the quantizer's own naming, or measured from the file The on-disk size of the weights. It is not the memory the model occupies: add the KV cache and the runtime overhead (SPEC §8.1). Preferred over the bits-per-weight formula wherever it exists, because it includes embeddings and any higher-precision tensors. Observed, not derived: this app records the figure and its source, and does not re-measure it. on disk	21.9 GB (estimate) ESTIMATEQwen3-VL 30B-A3B Instruct memory requirement at 32K contextArithmetic that also depends on a tunable assumption, listed below. required_gb = weights_gb + kv_cache_gb + overhead_gb required_gb = 17.31 + 3.22 + 1.38 = 21.92 GB weights_gb 17.310783264 GB — real file size kv_cache_gb 3.221225472 GB — kv formula overhead_gb 1.3848626611200001 GB — overhead fraction Real file size used instead of the params × bits formula (SPEC §8.1). 48 layers keep a full KV cache. Runtime buffers, activations, and allocator slack approximated as a fraction of weights.	82 of 89 (derived) DERIVEDmachines that run Qwen3-VL 30B-A3B Instruct at ≥ 15 tok/sArithmetic on observed inputs only. No tunable assumption enters it. machines = #{candidate rigs where the model loads AND clears the Plan's speed floor} 89 evaluated → 86 hold it in memory → 82 of those also reach ≥ 15 tok/s candidates_evaluated 89 machines — every complete rig the search could build from the priced catalog hold_it_in_memory 86 machines — required memory ≤ usable memory (SPEC §8.6) clear_the_speed_floor 82 machines — ≥ 15 tok/s The denominator is the machines this catalog can build today, not every machine that exists. A model with no shape or no sized quantization is counted on no machine at all, and says so rather than counting as zero.	SIZEABLE	read stale 3 d — past the 36 hour freshness window — Qwen3-VL 30B-A3B Instruct: the least recent read of its data from source (shape counted Oct 17, 8:06 AM ET from the tensor headers of https://huggingface.co/Qwen/Qwen3-VL-30B-A3B-Instruct; quantization sizes read Oct 17, 8:06 AM ET); never checked upstream.
Ministral 3 14B Instruct (2512)Mistral AI · dense · vision · 256K ctx max	13.5 (derived) DERIVEDMinistral 3 14B Instruct (2512) total parametersArithmetic on observed inputs only. No tunable assumption enters it. Σ params of language-model tensors — Cross-checked: tensor count 13.95B vs Hugging Face 13.95B (-2.01e-6%). 13.5 B, counted from the tensor headers of https://huggingface.co/mistralai/Ministral-3-14B-Instruct-2512 at 2026-10-17T13:05:38.498Z total_parameters 13.5 B — https://huggingface.co/mistralai/Ministral-3-14B-Instruct-2512 counted_at 2026-10-17T13:05:38.498Z — spec_sources.verified_at for total_params_b — when this figure was counted; a refresh that keeps it does not rewrite it confidence 0.95 — spec_sources.confidence — from the tensor headers The parameters of the language-model tensors only, counted from the tensor headers: a vision or audio tower and multi-token-prediction heads are excluded, so it can be below the whole-checkpoint count the formula's cross-check quotes. It sizes the weights only when no measured file size is stored for the chosen quantization; it is not what is read per token on a mixture-of-experts model. Shown to one decimal: the underlying tensor count is exact, but more digits would claim a precision the published figure does not have (SPEC §14). 12.8 (derived) DERIVEDMinistral 3 14B Instruct (2512) active parameters per tokenArithmetic on observed inputs only. No tunable assumption enters it. active − input embedding table — Per-token memory read used for decode bandwidth (SPEC §8.2). 12.8 B, counted from the tensor headers of https://huggingface.co/mistralai/Ministral-3-14B-Instruct-2512 at 2026-10-17T13:05:38.498Z active_parameters_per_token 12.8 B — https://huggingface.co/mistralai/Ministral-3-14B-Instruct-2512 counted_at 2026-10-17T13:05:38.498Z — spec_sources.verified_at for active_params_b — when this figure was counted; a refresh that keeps it does not rewrite it confidence 0.95 — spec_sources.confidence — from the tensor headers The parameters read to produce one token. It sets the decode speed; on a dense model it equals the total. Shown to one decimal: the underlying tensor count is exact, but more digits would claim a precision the published figure does not have (SPEC §14). active	IQ4_XS 7.43 GB (observed) OBSERVEDIQ4_XS file size on diskMeasured or published, passed through unchanged — not computed here. quantization file size — published at the source, passed through unchanged; nothing is computed here 7.43 GB, as recorded from https://huggingface.co/unsloth/Ministral-3-14B-Instruct-2512-GGUF/tree/main, file Ministral-3-14B-Instruct-2512-IQ4_XS.gguf quantization_file_size 7.43 GB — https://huggingface.co/unsloth/Ministral-3-14B-Instruct-2512-GGUF/tree/main, file Ministral-3-14B-Instruct-2512-IQ4_XS.gguf verified_at 2026-10-17T13:05:38.498Z — the row's own verification timestamp quantization IQ4_XS — gguf bits_per_weight 4.4023 bits — the quantizer's own naming, or measured from the file The on-disk size of the weights. It is not the memory the model occupies: add the KV cache and the runtime overhead (SPEC §8.1). Preferred over the bits-per-weight formula wherever it exists, because it includes embeddings and any higher-precision tensors. Observed, not derived: this app records the figure and its source, and does not re-measure it. on disk	13.4 GB (estimate) ESTIMATEMinistral 3 14B Instruct (2512) memory requirement at 32K contextArithmetic that also depends on a tunable assumption, listed below. required_gb = weights_gb + kv_cache_gb + overhead_gb required_gb = 7.43 + 5.37 + 0.59 = 13.4 GB weights_gb 7.43215584 GB — real file size kv_cache_gb 5.36870912 GB — kv formula overhead_gb 0.5945724672 GB — overhead fraction Real file size used instead of the params × bits formula (SPEC §8.1). 40 layers keep a full KV cache. Runtime buffers, activations, and allocator slack approximated as a fraction of weights.	82 of 89 (derived) DERIVEDmachines that run Ministral 3 14B Instruct (2512) at ≥ 15 tok/sArithmetic on observed inputs only. No tunable assumption enters it. machines = #{candidate rigs where the model loads AND clears the Plan's speed floor} 89 evaluated → 88 hold it in memory → 82 of those also reach ≥ 15 tok/s candidates_evaluated 89 machines — every complete rig the search could build from the priced catalog hold_it_in_memory 88 machines — required memory ≤ usable memory (SPEC §8.6) clear_the_speed_floor 82 machines — ≥ 15 tok/s The denominator is the machines this catalog can build today, not every machine that exists. A model with no shape or no sized quantization is counted on no machine at all, and says so rather than counting as zero.	SIZEABLE	read stale 3 d — past the 36 hour freshness window — Ministral 3 14B Instruct (2512): the least recent read of its data from source (shape counted Oct 17, 9:05 AM ET from the tensor headers of https://huggingface.co/mistralai/Ministral-3-14B-Instruct-2512; quantization sizes read Oct 17, 9:05 AM ET); never checked upstream.
Qwen3-VL 8B InstructAlibaba Qwen · dense · vision · 256K ctx max	8.2 (derived) DERIVEDQwen3-VL 8B Instruct total parametersArithmetic on observed inputs only. No tunable assumption enters it. Σ params of language-model tensors — Cross-checked: tensor count 8.77B vs Hugging Face 8.77B (0%). 8.2 B, counted from the tensor headers of https://huggingface.co/Qwen/Qwen3-VL-8B-Instruct at 2026-10-17T11:59:44.723Z total_parameters 8.2 B — https://huggingface.co/Qwen/Qwen3-VL-8B-Instruct counted_at 2026-10-17T11:59:44.723Z — spec_sources.verified_at for total_params_b — when this figure was counted; a refresh that keeps it does not rewrite it confidence 0.95 — spec_sources.confidence — from the tensor headers The parameters of the language-model tensors only, counted from the tensor headers: a vision or audio tower and multi-token-prediction heads are excluded, so it can be below the whole-checkpoint count the formula's cross-check quotes. It sizes the weights only when no measured file size is stored for the chosen quantization; it is not what is read per token on a mixture-of-experts model. Shown to one decimal: the underlying tensor count is exact, but more digits would claim a precision the published figure does not have (SPEC §14). 7.6 (derived) DERIVEDQwen3-VL 8B Instruct active parameters per tokenArithmetic on observed inputs only. No tunable assumption enters it. active − input embedding table — Per-token memory read used for decode bandwidth (SPEC §8.2). 7.6 B, counted from the tensor headers of https://huggingface.co/Qwen/Qwen3-VL-8B-Instruct at 2026-10-17T11:59:44.723Z active_parameters_per_token 7.6 B — https://huggingface.co/Qwen/Qwen3-VL-8B-Instruct counted_at 2026-10-17T11:59:44.723Z — spec_sources.verified_at for active_params_b — when this figure was counted; a refresh that keeps it does not rewrite it confidence 0.95 — spec_sources.confidence — from the tensor headers The parameters read to produce one token. It sets the decode speed; on a dense model it equals the total. Shown to one decimal: the underlying tensor count is exact, but more digits would claim a precision the published figure does not have (SPEC §14). active	IQ4_XS 4.58 GB (observed) OBSERVEDIQ4_XS file size on diskMeasured or published, passed through unchanged — not computed here. quantization file size — published at the source, passed through unchanged; nothing is computed here 4.58 GB, as recorded from https://huggingface.co/unsloth/Qwen3-VL-8B-Instruct-1M-GGUF/tree/main, file Qwen3-VL-8B-Instruct-1M-IQ4_XS.gguf quantization_file_size 4.58 GB — https://huggingface.co/unsloth/Qwen3-VL-8B-Instruct-1M-GGUF/tree/main, file Qwen3-VL-8B-Instruct-1M-IQ4_XS.gguf verified_at 2026-10-17T11:59:44.723Z — the row's own verification timestamp quantization IQ4_XS — gguf bits_per_weight 4.4746 bits — the quantizer's own naming, or measured from the file The on-disk size of the weights. It is not the memory the model occupies: add the KV cache and the runtime overhead (SPEC §8.1). Preferred over the bits-per-weight formula wherever it exists, because it includes embeddings and any higher-precision tensors. Observed, not derived: this app records the figure and its source, and does not re-measure it. on disk	9.78 GB (estimate) ESTIMATEQwen3-VL 8B Instruct memory requirement at 32K contextArithmetic that also depends on a tunable assumption, listed below. required_gb = weights_gb + kv_cache_gb + overhead_gb required_gb = 4.58 + 4.83 + 0.37 = 9.78 GB weights_gb 4.58128896 GB — real file size kv_cache_gb 4.831838208 GB — kv formula overhead_gb 0.36650311680000003 GB — overhead fraction Real file size used instead of the params × bits formula (SPEC §8.1). 36 layers keep a full KV cache. Runtime buffers, activations, and allocator slack approximated as a fraction of weights.	82 of 89 (derived) DERIVEDmachines that run Qwen3-VL 8B Instruct at ≥ 15 tok/sArithmetic on observed inputs only. No tunable assumption enters it. machines = #{candidate rigs where the model loads AND clears the Plan's speed floor} 89 evaluated → 89 hold it in memory → 82 of those also reach ≥ 15 tok/s candidates_evaluated 89 machines — every complete rig the search could build from the priced catalog hold_it_in_memory 89 machines — required memory ≤ usable memory (SPEC §8.6) clear_the_speed_floor 82 machines — ≥ 15 tok/s The denominator is the machines this catalog can build today, not every machine that exists. A model with no shape or no sized quantization is counted on no machine at all, and says so rather than counting as zero.	SIZEABLE	read stale 3 d — past the 36 hour freshness window — Qwen3-VL 8B Instruct: the least recent read of its data from source (shape counted Oct 17, 7:59 AM ET from the tensor headers of https://huggingface.co/Qwen/Qwen3-VL-8B-Instruct; quantization sizes read Oct 17, 7:59 AM ET); never checked upstream.
InternVL3.5 8BShanghai AI Lab (OpenGVLab) · dense · vision · 40K ctx max	8.2 (derived) DERIVEDInternVL3.5 8B total parametersArithmetic on observed inputs only. No tunable assumption enters it. Σ params of language-model tensors — Cross-checked: tensor count 8.53B vs Hugging Face 8.53B (0%). 8.2 B, counted from the tensor headers of https://huggingface.co/OpenGVLab/InternVL3_5-8B at 2026-10-17T12:32:14.805Z total_parameters 8.2 B — https://huggingface.co/OpenGVLab/InternVL3_5-8B counted_at 2026-10-17T12:32:14.805Z — spec_sources.verified_at for total_params_b — when this figure was counted; a refresh that keeps it does not rewrite it confidence 0.95 — spec_sources.confidence — from the tensor headers The parameters of the language-model tensors only, counted from the tensor headers: a vision or audio tower and multi-token-prediction heads are excluded, so it can be below the whole-checkpoint count the formula's cross-check quotes. It sizes the weights only when no measured file size is stored for the chosen quantization; it is not what is read per token on a mixture-of-experts model. Shown to one decimal: the underlying tensor count is exact, but more digits would claim a precision the published figure does not have (SPEC §14). 7.6 (derived) DERIVEDInternVL3.5 8B active parameters per tokenArithmetic on observed inputs only. No tunable assumption enters it. active − input embedding table — Per-token memory read used for decode bandwidth (SPEC §8.2). 7.6 B, counted from the tensor headers of https://huggingface.co/OpenGVLab/InternVL3_5-8B at 2026-10-17T12:32:14.805Z active_parameters_per_token 7.6 B — https://huggingface.co/OpenGVLab/InternVL3_5-8B counted_at 2026-10-17T12:32:14.805Z — spec_sources.verified_at for active_params_b — when this figure was counted; a refresh that keeps it does not rewrite it confidence 0.95 — spec_sources.confidence — from the tensor headers The parameters read to produce one token. It sets the decode speed; on a dense model it equals the total. Shown to one decimal: the underlying tensor count is exact, but more digits would claim a precision the published figure does not have (SPEC §14). active	IQ4_XS 4.56 GB (observed) OBSERVEDIQ4_XS file size on diskMeasured or published, passed through unchanged — not computed here. quantization file size — published at the source, passed through unchanged; nothing is computed here 4.56 GB, as recorded from https://huggingface.co/bartowski/OpenGVLab_InternVL3_5-8B-GGUF/tree/main, file OpenGVLab_InternVL3_5-8B-IQ4_XS.gguf quantization_file_size 4.56 GB — https://huggingface.co/bartowski/OpenGVLab_InternVL3_5-8B-GGUF/tree/main, file OpenGVLab_InternVL3_5-8B-IQ4_XS.gguf verified_at 2026-10-17T12:32:14.805Z — the row's own verification timestamp quantization IQ4_XS — gguf bits_per_weight 4.4374 bits — the quantizer's own naming, or measured from the file The on-disk size of the weights. It is not the memory the model occupies: add the KV cache and the runtime overhead (SPEC §8.1). Preferred over the bits-per-weight formula wherever it exists, because it includes embeddings and any higher-precision tensors. Observed, not derived: this app records the figure and its source, and does not re-measure it. on disk	9.76 GB (estimate) ESTIMATEInternVL3.5 8B memory requirement at 32K contextArithmetic that also depends on a tunable assumption, listed below. required_gb = weights_gb + kv_cache_gb + overhead_gb required_gb = 4.56 + 4.83 + 0.36 = 9.76 GB weights_gb 4.561836224 GB — real file size kv_cache_gb 4.831838208 GB — kv formula overhead_gb 0.36494689792 GB — overhead fraction Real file size used instead of the params × bits formula (SPEC §8.1). 36 layers keep a full KV cache. Runtime buffers, activations, and allocator slack approximated as a fraction of weights.	82 of 89 (derived) DERIVEDmachines that run InternVL3.5 8B at ≥ 15 tok/sArithmetic on observed inputs only. No tunable assumption enters it. machines = #{candidate rigs where the model loads AND clears the Plan's speed floor} 89 evaluated → 89 hold it in memory → 82 of those also reach ≥ 15 tok/s candidates_evaluated 89 machines — every complete rig the search could build from the priced catalog hold_it_in_memory 89 machines — required memory ≤ usable memory (SPEC §8.6) clear_the_speed_floor 82 machines — ≥ 15 tok/s The denominator is the machines this catalog can build today, not every machine that exists. A model with no shape or no sized quantization is counted on no machine at all, and says so rather than counting as zero.	SIZEABLE	read stale 3 d — past the 36 hour freshness window — InternVL3.5 8B: the least recent read of its data from source (shape counted Oct 17, 8:32 AM ET from the tensor headers of https://huggingface.co/OpenGVLab/InternVL3_5-8B; quantization sizes read Oct 17, 8:32 AM ET); never checked upstream.
gpt-oss-120bOpenAI · MoE · reasoning · 128K ctx max	116.8 (derived) DERIVEDgpt-oss-120b total parametersArithmetic on observed inputs only. No tunable assumption enters it. Σ params of language-model tensors — Cross-checked: tensor count 116.83B vs Hugging Face 116.83B (0%). 116.8 B, counted from the tensor headers of https://huggingface.co/openai/gpt-oss-120b at 2026-10-17T10:32:50.257Z total_parameters 116.8 B — https://huggingface.co/openai/gpt-oss-120b counted_at 2026-10-17T10:32:50.257Z — spec_sources.verified_at for total_params_b — when this figure was counted; a refresh that keeps it does not rewrite it confidence 0.95 — spec_sources.confidence — from the tensor headers The parameters of the language-model tensors only, counted from the tensor headers: a vision or audio tower and multi-token-prediction heads are excluded, so it can be below the whole-checkpoint count the formula's cross-check quotes. It sizes the weights only when no measured file size is stored for the chosen quantization; it is not what is read per token on a mixture-of-experts model. Shown to one decimal: the underlying tensor count is exact, but more digits would claim a precision the published figure does not have (SPEC §14). 5.1 (derived) DERIVEDgpt-oss-120b active parameters per tokenArithmetic on observed inputs only. No tunable assumption enters it. active − input embedding table — Per-token memory read used for decode bandwidth (SPEC §8.2). 5.1 B, counted from the tensor headers of https://huggingface.co/openai/gpt-oss-120b at 2026-10-17T10:32:50.257Z active_parameters_per_token 5.1 B — https://huggingface.co/openai/gpt-oss-120b counted_at 2026-10-17T10:32:50.257Z — spec_sources.verified_at for active_params_b — when this figure was counted; a refresh that keeps it does not rewrite it confidence 0.95 — spec_sources.confidence — from the tensor headers The parameters read to produce one token. It sets the decode speed; on a dense model it equals the total. Shown to one decimal: the underlying tensor count is exact, but more digits would claim a precision the published figure does not have (SPEC §14). active	BF16 65.4 GB (observed) OBSERVEDBF16 file size on diskMeasured or published, passed through unchanged — not computed here. quantization file size — published at the source, passed through unchanged; nothing is computed here 65.37 GB, as recorded from https://huggingface.co/bartowski/openai_gpt-oss-120b-GGUF/tree/main, file 2 files quantization_file_size 65.37 GB — https://huggingface.co/bartowski/openai_gpt-oss-120b-GGUF/tree/main, file 2 files verified_at 2026-10-17T10:32:50.257Z — the row's own verification timestamp quantization BF16 — gguf bits_per_weight 4.4762 bits — the quantizer's own naming, or measured from the file The on-disk size of the weights. It is not the memory the model occupies: add the KV cache and the runtime overhead (SPEC §8.1). Preferred over the bits-per-weight formula wherever it exists, because it includes embeddings and any higher-precision tensors. Observed, not derived: this app records the figure and its source, and does not re-measure it. on disk	71.8 GB (estimate) ESTIMATEgpt-oss-120b memory requirement at 32K contextArithmetic that also depends on a tunable assumption, listed below. required_gb = weights_gb + kv_cache_gb + overhead_gb required_gb = 65.37 + 1.21 + 5.23 = 71.81 GB weights_gb 65.369016768 GB — real file size kv_cache_gb 1.212678144 GB — kv formula overhead_gb 5.22952134144 GB — overhead fraction Real file size used instead of the params × bits formula (SPEC §8.1). 18 layers keep a full KV cache, 18 sliding-window layers are capped at 128 tokens. Runtime buffers, activations, and allocator slack approximated as a fraction of weights.	74 of 89 (derived) DERIVEDmachines that run gpt-oss-120b at ≥ 15 tok/sArithmetic on observed inputs only. No tunable assumption enters it. machines = #{candidate rigs where the model loads AND clears the Plan's speed floor} 89 evaluated → 74 hold it in memory → 74 of those also reach ≥ 15 tok/s candidates_evaluated 89 machines — every complete rig the search could build from the priced catalog hold_it_in_memory 74 machines — required memory ≤ usable memory (SPEC §8.6) clear_the_speed_floor 74 machines — ≥ 15 tok/s The denominator is the machines this catalog can build today, not every machine that exists. A model with no shape or no sized quantization is counted on no machine at all, and says so rather than counting as zero.	SIZEABLE	read stale 3 d — past the 36 hour freshness window — gpt-oss-120b: the least recent read of its data from source (shape counted Oct 17, 6:32 AM ET from the tensor headers of https://huggingface.co/openai/gpt-oss-120b; quantization sizes read Oct 17, 6:32 AM ET); never checked upstream.
Mistral Small 4 119B (2603)Mistral AI · MoE · vision, reasoning · 1M ctx max	119.0 (derived) DERIVEDMistral Small 4 119B (2603) total parametersArithmetic on observed inputs only. No tunable assumption enters it. Σ params of language-model tensors — Cross-checked: tensor count 119.4B vs Hugging Face 119.4B (-7.96e-6%). 119.0 B, counted from the tensor headers of https://huggingface.co/mistralai/Mistral-Small-4-119B-2603 at 2026-10-17T13:07:26.681Z total_parameters 119.0 B — https://huggingface.co/mistralai/Mistral-Small-4-119B-2603 counted_at 2026-10-17T13:07:26.681Z — spec_sources.verified_at for total_params_b — when this figure was counted; a refresh that keeps it does not rewrite it confidence 0.95 — spec_sources.confidence — from the tensor headers The parameters of the language-model tensors only, counted from the tensor headers: a vision or audio tower and multi-token-prediction heads are excluded, so it can be below the whole-checkpoint count the formula's cross-check quotes. It sizes the weights only when no measured file size is stored for the chosen quantization; it is not what is read per token on a mixture-of-experts model. Shown to one decimal: the underlying tensor count is exact, but more digits would claim a precision the published figure does not have (SPEC §14). 6.1 (derived) DERIVEDMistral Small 4 119B (2603) active parameters per tokenArithmetic on observed inputs only. No tunable assumption enters it. active − input embedding table — Per-token memory read used for decode bandwidth (SPEC §8.2). 6.1 B, counted from the tensor headers of https://huggingface.co/mistralai/Mistral-Small-4-119B-2603 at 2026-10-17T13:07:26.681Z active_parameters_per_token 6.1 B — https://huggingface.co/mistralai/Mistral-Small-4-119B-2603 counted_at 2026-10-17T13:07:26.681Z — spec_sources.verified_at for active_params_b — when this figure was counted; a refresh that keeps it does not rewrite it confidence 0.95 — spec_sources.confidence — from the tensor headers The parameters read to produce one token. It sets the decode speed; on a dense model it equals the total. Shown to one decimal: the underlying tensor count is exact, but more digits would claim a precision the published figure does not have (SPEC §14). active	IQ4_NL 67.5 GB (observed) OBSERVEDIQ4_NL file size on diskMeasured or published, passed through unchanged — not computed here. quantization file size — published at the source, passed through unchanged; nothing is computed here 67.46 GB, as recorded from https://huggingface.co/bartowski/mistralai_Mistral-Small-4-119B-2603-GGUF/tree/main, file 2 files quantization_file_size 67.46 GB — https://huggingface.co/bartowski/mistralai_Mistral-Small-4-119B-2603-GGUF/tree/main, file 2 files verified_at 2026-10-17T13:07:26.681Z — the row's own verification timestamp quantization IQ4_NL — gguf bits_per_weight 4.5362 bits — the quantizer's own naming, or measured from the file The on-disk size of the weights. It is not the memory the model occupies: add the KV cache and the runtime overhead (SPEC §8.1). Preferred over the bits-per-weight formula wherever it exists, because it includes embeddings and any higher-precision tensors. Observed, not derived: this app records the figure and its source, and does not re-measure it. on disk	73.6 GB (estimate) ESTIMATEMistral Small 4 119B (2603) memory requirement at 32K contextArithmetic that also depends on a tunable assumption, listed below. required_gb = weights_gb + kv_cache_gb + overhead_gb required_gb = 67.46 + 0.75 + 5.4 = 73.61 GB weights_gb 67.460181472 GB — real file size kv_cache_gb 0.75497472 GB — kv formula overhead_gb 5.39681451776 GB — overhead fraction Real file size used instead of the params × bits formula (SPEC §8.1). 36 layers keep a full KV cache. Runtime buffers, activations, and allocator slack approximated as a fraction of weights.	70 of 89 (derived) DERIVEDmachines that run Mistral Small 4 119B (2603) at ≥ 15 tok/sArithmetic on observed inputs only. No tunable assumption enters it. machines = #{candidate rigs where the model loads AND clears the Plan's speed floor} 89 evaluated → 70 hold it in memory → 70 of those also reach ≥ 15 tok/s candidates_evaluated 89 machines — every complete rig the search could build from the priced catalog hold_it_in_memory 70 machines — required memory ≤ usable memory (SPEC §8.6) clear_the_speed_floor 70 machines — ≥ 15 tok/s The denominator is the machines this catalog can build today, not every machine that exists. A model with no shape or no sized quantization is counted on no machine at all, and says so rather than counting as zero.	SIZEABLE	read stale 3 d — past the 36 hour freshness window — Mistral Small 4 119B (2603): the least recent read of its data from source (shape counted Oct 17, 9:07 AM ET from the tensor headers of https://huggingface.co/mistralai/Mistral-Small-4-119B-2603; quantization sizes read Oct 17, 9:07 AM ET); never checked upstream.
Muse Glimmer 30BMeta (meta-models) · dense · vision, coding, reasoning · 128K ctx max	27.9 (derived) DERIVEDMuse Glimmer 30B total parametersArithmetic on observed inputs only. No tunable assumption enters it. Σ params of language-model tensors — Cross-checked: tensor count 29.78B vs Hugging Face 29.78B (0%). 27.9 B, counted from the tensor headers of https://huggingface.co/meta-models/Muse-Glimmer-30B at 2026-10-17T13:03:15.445Z total_parameters 27.9 B — https://huggingface.co/meta-models/Muse-Glimmer-30B counted_at 2026-10-17T13:03:15.445Z — spec_sources.verified_at for total_params_b — when this figure was counted; a refresh that keeps it does not rewrite it confidence 0.95 — spec_sources.confidence — from the tensor headers The parameters of the language-model tensors only, counted from the tensor headers: a vision or audio tower and multi-token-prediction heads are excluded, so it can be below the whole-checkpoint count the formula's cross-check quotes. It sizes the weights only when no measured file size is stored for the chosen quantization; it is not what is read per token on a mixture-of-experts model. Shown to one decimal: the underlying tensor count is exact, but more digits would claim a precision the published figure does not have (SPEC §14). 26.6 (derived) DERIVEDMuse Glimmer 30B active parameters per tokenArithmetic on observed inputs only. No tunable assumption enters it. active − input embedding table — Per-token memory read used for decode bandwidth (SPEC §8.2). 26.6 B, counted from the tensor headers of https://huggingface.co/meta-models/Muse-Glimmer-30B at 2026-10-17T13:03:15.445Z active_parameters_per_token 26.6 B — https://huggingface.co/meta-models/Muse-Glimmer-30B counted_at 2026-10-17T13:03:15.445Z — spec_sources.verified_at for active_params_b — when this figure was counted; a refresh that keeps it does not rewrite it confidence 0.95 — spec_sources.confidence — from the tensor headers The parameters read to produce one token. It sets the decode speed; on a dense model it equals the total. Shown to one decimal: the underlying tensor count is exact, but more digits would claim a precision the published figure does not have (SPEC §14). active	UD-Q4_K_XL 15.9 GB (observed) OBSERVEDUD-Q4_K_XL file size on diskMeasured or published, passed through unchanged — not computed here. quantization file size — published at the source, passed through unchanged; nothing is computed here 15.88 GB, as recorded from https://huggingface.co/unsloth/Muse-Glimmer-30B-GGUF/tree/main, file Muse-Glimmer-30B-UD-Q4_K_XL.gguf quantization_file_size 15.88 GB — https://huggingface.co/unsloth/Muse-Glimmer-30B-GGUF/tree/main, file Muse-Glimmer-30B-UD-Q4_K_XL.gguf verified_at 2026-10-17T13:03:15.445Z — the row's own verification timestamp quantization UD-Q4_K_XL — gguf bits_per_weight 4.549 bits — the quantizer's own naming, or measured from the file The on-disk size of the weights. It is not the memory the model occupies: add the KV cache and the runtime overhead (SPEC §8.1). Preferred over the bits-per-weight formula wherever it exists, because it includes embeddings and any higher-precision tensors. Observed, not derived: this app records the figure and its source, and does not re-measure it. on disk	17.7 GB (estimate) ESTIMATEMuse Glimmer 30B memory requirement at 32K contextArithmetic that also depends on a tunable assumption, listed below. required_gb = weights_gb + kv_cache_gb + overhead_gb required_gb = 15.88 + 0.52 + 1.27 = 17.67 GB weights_gb 15.878222368 GB — real file size kv_cache_gb 0.517996544 GB — kv formula overhead_gb 1.27025778944 GB — overhead fraction Real file size used instead of the params × bits formula (SPEC §8.1). 13 layers keep a full KV cache, 39 sliding-window layers are capped at 2048 tokens. Runtime buffers, activations, and allocator slack approximated as a fraction of weights.	61 of 89 (derived) DERIVEDmachines that run Muse Glimmer 30B at ≥ 15 tok/sArithmetic on observed inputs only. No tunable assumption enters it. machines = #{candidate rigs where the model loads AND clears the Plan's speed floor} 89 evaluated → 88 hold it in memory → 61 of those also reach ≥ 15 tok/s candidates_evaluated 89 machines — every complete rig the search could build from the priced catalog hold_it_in_memory 88 machines — required memory ≤ usable memory (SPEC §8.6) clear_the_speed_floor 61 machines — ≥ 15 tok/s The denominator is the machines this catalog can build today, not every machine that exists. A model with no shape or no sized quantization is counted on no machine at all, and says so rather than counting as zero.	SIZEABLE	read stale 3 d — past the 36 hour freshness window — Muse Glimmer 30B: the least recent read of its data from source (shape counted Oct 17, 9:03 AM ET from the tensor headers of https://huggingface.co/meta-models/Muse-Glimmer-30B; quantization sizes read Oct 17, 9:03 AM ET); never checked upstream.
gemma-4-31B-itGoogle DeepMind · dense · vision, reasoning, coding · 256K ctx max	30.7 (derived) DERIVEDgemma-4-31B-it total parametersArithmetic on observed inputs only. No tunable assumption enters it. Σ params of language-model tensors — Cross-checked: tensor count 31.27B vs Hugging Face 31.27B (0%). 30.7 B, counted from the tensor headers of https://huggingface.co/google/gemma-4-31B-it at 2026-10-17T11:27:52.958Z total_parameters 30.7 B — https://huggingface.co/google/gemma-4-31B-it counted_at 2026-10-17T11:27:52.958Z — spec_sources.verified_at for total_params_b — when this figure was counted; a refresh that keeps it does not rewrite it confidence 0.95 — spec_sources.confidence — from the tensor headers The parameters of the language-model tensors only, counted from the tensor headers: a vision or audio tower and multi-token-prediction heads are excluded, so it can be below the whole-checkpoint count the formula's cross-check quotes. It sizes the weights only when no measured file size is stored for the chosen quantization; it is not what is read per token on a mixture-of-experts model. Shown to one decimal: the underlying tensor count is exact, but more digits would claim a precision the published figure does not have (SPEC §14). 29.3 (derived) DERIVEDgemma-4-31B-it active parameters per tokenArithmetic on observed inputs only. No tunable assumption enters it. active − input embedding table — Per-token memory read used for decode bandwidth (SPEC §8.2). 29.3 B, counted from the tensor headers of https://huggingface.co/google/gemma-4-31B-it at 2026-10-17T11:27:52.958Z active_parameters_per_token 29.3 B — https://huggingface.co/google/gemma-4-31B-it counted_at 2026-10-17T11:27:52.958Z — spec_sources.verified_at for active_params_b — when this figure was counted; a refresh that keeps it does not rewrite it confidence 0.95 — spec_sources.confidence — from the tensor headers The parameters read to produce one token. It sets the decode speed; on a dense model it equals the total. Shown to one decimal: the underlying tensor count is exact, but more digits would claim a precision the published figure does not have (SPEC §14). active	IQ4_NL 17.3 GB (observed) OBSERVEDIQ4_NL file size on diskMeasured or published, passed through unchanged — not computed here. quantization file size — published at the source, passed through unchanged; nothing is computed here 17.29 GB, as recorded from https://huggingface.co/unsloth/gemma-4-31B-it-GGUF/tree/main, file gemma-4-31B-it-IQ4_NL.gguf quantization_file_size 17.29 GB — https://huggingface.co/unsloth/gemma-4-31B-it-GGUF/tree/main, file gemma-4-31B-it-IQ4_NL.gguf verified_at 2026-10-17T11:27:52.958Z — the row's own verification timestamp quantization IQ4_NL — gguf bits_per_weight 4.5053 bits — the quantizer's own naming, or measured from the file The on-disk size of the weights. It is not the memory the model occupies: add the KV cache and the runtime overhead (SPEC §8.1). Preferred over the bits-per-weight formula wherever it exists, because it includes embeddings and any higher-precision tensors. Observed, not derived: this app records the figure and its source, and does not re-measure it. on disk	20.9 GB (estimate) ESTIMATEgemma-4-31B-it memory requirement at 32K contextArithmetic that also depends on a tunable assumption, listed below. required_gb = weights_gb + kv_cache_gb + overhead_gb required_gb = 17.29 + 2.18 + 1.38 = 20.85 GB weights_gb 17.28767072 GB — real file size kv_cache_gb 2.18103808 GB — kv formula overhead_gb 1.3830136576 GB — overhead fraction Real file size used instead of the params × bits formula (SPEC §8.1). 10 layers keep a full KV cache, 50 sliding-window layers are capped at 1024 tokens. Runtime buffers, activations, and allocator slack approximated as a fraction of weights.	57 of 89 (derived) DERIVEDmachines that run gemma-4-31B-it at ≥ 15 tok/sArithmetic on observed inputs only. No tunable assumption enters it. machines = #{candidate rigs where the model loads AND clears the Plan's speed floor} 89 evaluated → 86 hold it in memory → 57 of those also reach ≥ 15 tok/s candidates_evaluated 89 machines — every complete rig the search could build from the priced catalog hold_it_in_memory 86 machines — required memory ≤ usable memory (SPEC §8.6) clear_the_speed_floor 57 machines — ≥ 15 tok/s The denominator is the machines this catalog can build today, not every machine that exists. A model with no shape or no sized quantization is counted on no machine at all, and says so rather than counting as zero.	SIZEABLE	read stale 3 d — past the 36 hour freshness window — gemma-4-31B-it: the least recent read of its data from source (shape counted Oct 17, 7:27 AM ET from the tensor headers of https://huggingface.co/google/gemma-4-31B-it; quantization sizes read Oct 17, 7:27 AM ET); never checked upstream.
Devstral-Small-2-24B-Instruct-2512Mistral AI · dense · coding · 384K ctx max	23.6 (derived) DERIVEDDevstral-Small-2-24B-Instruct-2512 total parametersArithmetic on observed inputs only. No tunable assumption enters it. Σ params of language-model tensors — Cross-checked: tensor count 24.01B vs Hugging Face 24.01B (-1.17e-6%). 23.6 B, counted from the tensor headers of https://huggingface.co/mistralai/Devstral-Small-2-24B-Instruct-2512 at 2026-10-17T11:24:45.213Z total_parameters 23.6 B — https://huggingface.co/mistralai/Devstral-Small-2-24B-Instruct-2512 counted_at 2026-10-17T11:24:45.213Z — spec_sources.verified_at for total_params_b — when this figure was counted; a refresh that keeps it does not rewrite it confidence 0.95 — spec_sources.confidence — from the tensor headers The parameters of the language-model tensors only, counted from the tensor headers: a vision or audio tower and multi-token-prediction heads are excluded, so it can be below the whole-checkpoint count the formula's cross-check quotes. It sizes the weights only when no measured file size is stored for the chosen quantization; it is not what is read per token on a mixture-of-experts model. Shown to one decimal: the underlying tensor count is exact, but more digits would claim a precision the published figure does not have (SPEC §14). 22.9 (derived) DERIVEDDevstral-Small-2-24B-Instruct-2512 active parameters per tokenArithmetic on observed inputs only. No tunable assumption enters it. active − input embedding table — Per-token memory read used for decode bandwidth (SPEC §8.2). 22.9 B, counted from the tensor headers of https://huggingface.co/mistralai/Devstral-Small-2-24B-Instruct-2512 at 2026-10-17T11:24:45.213Z active_parameters_per_token 22.9 B — https://huggingface.co/mistralai/Devstral-Small-2-24B-Instruct-2512 counted_at 2026-10-17T11:24:45.213Z — spec_sources.verified_at for active_params_b — when this figure was counted; a refresh that keeps it does not rewrite it confidence 0.95 — spec_sources.confidence — from the tensor headers The parameters read to produce one token. It sets the decode speed; on a dense model it equals the total. Shown to one decimal: the underlying tensor count is exact, but more digits would claim a precision the published figure does not have (SPEC §14). active	IQ4_NL 13.5 GB (observed) OBSERVEDIQ4_NL file size on diskMeasured or published, passed through unchanged — not computed here. quantization file size — published at the source, passed through unchanged; nothing is computed here 13.47 GB, as recorded from https://huggingface.co/bartowski/mistralai_Devstral-Small-2-24B-Instruct-2512-GGUF/tree/main, file mistralai_Devstral-Small-2-24B-Instruct-2512-IQ4_NL.gguf quantization_file_size 13.47 GB — https://huggingface.co/bartowski/mistralai_Devstral-Small-2-24B-Instruct-2512-GGUF/tree/main, file mistralai_Devstral-Small-2-24B-Instruct-2512-IQ4_NL.gguf verified_at 2026-10-17T11:24:45.213Z — the row's own verification timestamp quantization IQ4_NL — gguf bits_per_weight 4.571 bits — the quantizer's own naming, or measured from the file The on-disk size of the weights. It is not the memory the model occupies: add the KV cache and the runtime overhead (SPEC §8.1). Preferred over the bits-per-weight formula wherever it exists, because it includes embeddings and any higher-precision tensors. Observed, not derived: this app records the figure and its source, and does not re-measure it. on disk	19.9 GB (estimate) ESTIMATEDevstral-Small-2-24B-Instruct-2512 memory requirement at 32K contextArithmetic that also depends on a tunable assumption, listed below. required_gb = weights_gb + kv_cache_gb + overhead_gb required_gb = 13.47 + 5.37 + 1.08 = 19.91 GB weights_gb 13.468543872 GB — real file size kv_cache_gb 5.36870912 GB — kv formula overhead_gb 1.07748350976 GB — overhead fraction Real file size used instead of the params × bits formula (SPEC §8.1). 40 layers keep a full KV cache. Runtime buffers, activations, and allocator slack approximated as a fraction of weights.	57 of 89 (derived) DERIVEDmachines that run Devstral-Small-2-24B-Instruct-2512 at ≥ 15 tok/sArithmetic on observed inputs only. No tunable assumption enters it. machines = #{candidate rigs where the model loads AND clears the Plan's speed floor} 89 evaluated → 86 hold it in memory → 57 of those also reach ≥ 15 tok/s candidates_evaluated 89 machines — every complete rig the search could build from the priced catalog hold_it_in_memory 86 machines — required memory ≤ usable memory (SPEC §8.6) clear_the_speed_floor 57 machines — ≥ 15 tok/s The denominator is the machines this catalog can build today, not every machine that exists. A model with no shape or no sized quantization is counted on no machine at all, and says so rather than counting as zero.	SIZEABLE	read stale 3 d — past the 36 hour freshness window — Devstral-Small-2-24B-Instruct-2512: the least recent read of its data from source (shape counted Oct 17, 7:24 AM ET from the tensor headers of https://huggingface.co/mistralai/Devstral-Small-2-24B-Instruct-2512; quantization sizes read Oct 17, 7:24 AM ET); never checked upstream.
Qwen3.6-27BAlibaba Qwen · dense · coding, reasoning, vision · 256K ctx max	26.9 (derived) DERIVEDQwen3.6-27B total parametersArithmetic on observed inputs only. No tunable assumption enters it. Σ params of language-model tensors — Cross-checked: tensor count 27.78B vs Hugging Face 27.78B (0%). 26.9 B, counted from the tensor headers of https://huggingface.co/Qwen/Qwen3.6-27B at 2026-10-18T07:30:46.328Z total_parameters 26.9 B — https://huggingface.co/Qwen/Qwen3.6-27B counted_at 2026-10-18T07:30:46.328Z — spec_sources.verified_at for total_params_b — when this figure was counted; a refresh that keeps it does not rewrite it confidence 0.95 — spec_sources.confidence — from the tensor headers The parameters of the language-model tensors only, counted from the tensor headers: a vision or audio tower and multi-token-prediction heads are excluded, so it can be below the whole-checkpoint count the formula's cross-check quotes. It sizes the weights only when no measured file size is stored for the chosen quantization; it is not what is read per token on a mixture-of-experts model. Shown to one decimal: the underlying tensor count is exact, but more digits would claim a precision the published figure does not have (SPEC §14). 25.6 (derived) DERIVEDQwen3.6-27B active parameters per tokenArithmetic on observed inputs only. No tunable assumption enters it. active − input embedding table — Per-token memory read used for decode bandwidth (SPEC §8.2). 25.6 B, counted from the tensor headers of https://huggingface.co/Qwen/Qwen3.6-27B at 2026-10-18T07:30:46.328Z active_parameters_per_token 25.6 B — https://huggingface.co/Qwen/Qwen3.6-27B counted_at 2026-10-18T07:30:46.328Z — spec_sources.verified_at for active_params_b — when this figure was counted; a refresh that keeps it does not rewrite it confidence 0.95 — spec_sources.confidence — from the tensor headers The parameters read to produce one token. It sets the decode speed; on a dense model it equals the total. Shown to one decimal: the underlying tensor count is exact, but more digits would claim a precision the published figure does not have (SPEC §14). active	IQ4_XS 15.4 GB (observed) OBSERVEDIQ4_XS file size on diskMeasured or published, passed through unchanged — not computed here. quantization file size — published at the source, passed through unchanged; nothing is computed here 15.44 GB, as recorded from https://huggingface.co/unsloth/Qwen3.6-27B-GGUF/tree/main, file Qwen3.6-27B-IQ4_XS.gguf quantization_file_size 15.44 GB — https://huggingface.co/unsloth/Qwen3.6-27B-GGUF/tree/main, file Qwen3.6-27B-IQ4_XS.gguf verified_at 2026-10-18T07:30:46.328Z — the row's own verification timestamp quantization IQ4_XS — gguf bits_per_weight 4.5925 bits — the quantizer's own naming, or measured from the file The on-disk size of the weights. It is not the memory the model occupies: add the KV cache and the runtime overhead (SPEC §8.1). Preferred over the bits-per-weight formula wherever it exists, because it includes embeddings and any higher-precision tensors. Observed, not derived: this app records the figure and its source, and does not re-measure it. on disk	18.8 GB (estimate) ESTIMATEQwen3.6-27B memory requirement at 32K contextArithmetic that also depends on a tunable assumption, listed below. required_gb = weights_gb + kv_cache_gb + overhead_gb required_gb = 15.44 + 2.15 + 1.24 = 18.82 GB weights_gb 15.440005344 GB — real file size kv_cache_gb 2.147483648 GB — kv formula overhead_gb 1.23520042752 GB — overhead fraction Real file size used instead of the params × bits formula (SPEC §8.1). 16 layers keep a full KV cache, 48 linear-attention/SSM layers keep only a constant-size state (treated as negligible). Runtime buffers, activations, and allocator slack approximated as a fraction of weights.	57 of 89 (derived) DERIVEDmachines that run Qwen3.6-27B at ≥ 15 tok/sArithmetic on observed inputs only. No tunable assumption enters it. machines = #{candidate rigs where the model loads AND clears the Plan's speed floor} 89 evaluated → 86 hold it in memory → 57 of those also reach ≥ 15 tok/s candidates_evaluated 89 machines — every complete rig the search could build from the priced catalog hold_it_in_memory 86 machines — required memory ≤ usable memory (SPEC §8.6) clear_the_speed_floor 57 machines — ≥ 15 tok/s The denominator is the machines this catalog can build today, not every machine that exists. A model with no shape or no sized quantization is counted on no machine at all, and says so rather than counting as zero.	SIZEABLE	read stale 2 d — past the 36 hour freshness window — Qwen3.6-27B: the least recent read of its data from source (shape counted Oct 18, 3:30 AM ET from the tensor headers of https://huggingface.co/Qwen/Qwen3.6-27B; quantization sizes read Oct 18, 3:30 AM ET); no upstream check after it.
Qwen3.8-27BAlibaba Qwen · dense · coding, reasoning, vision · 256K ctx max	26.9 (derived) DERIVEDQwen3.8-27B total parametersArithmetic on observed inputs only. No tunable assumption enters it. Σ params of language-model tensors — Cross-checked: tensor count 27.78B vs Hugging Face 27.78B (0%). 26.9 B, counted from the tensor headers of https://huggingface.co/Qwen/Qwen3.8-27B at 2026-10-18T07:31:57.966Z total_parameters 26.9 B — https://huggingface.co/Qwen/Qwen3.8-27B counted_at 2026-10-18T07:31:57.966Z — spec_sources.verified_at for total_params_b — when this figure was counted; a refresh that keeps it does not rewrite it confidence 0.95 — spec_sources.confidence — from the tensor headers The parameters of the language-model tensors only, counted from the tensor headers: a vision or audio tower and multi-token-prediction heads are excluded, so it can be below the whole-checkpoint count the formula's cross-check quotes. It sizes the weights only when no measured file size is stored for the chosen quantization; it is not what is read per token on a mixture-of-experts model. Shown to one decimal: the underlying tensor count is exact, but more digits would claim a precision the published figure does not have (SPEC §14). 25.6 (derived) DERIVEDQwen3.8-27B active parameters per tokenArithmetic on observed inputs only. No tunable assumption enters it. active − input embedding table — Per-token memory read used for decode bandwidth (SPEC §8.2). 25.6 B, counted from the tensor headers of https://huggingface.co/Qwen/Qwen3.8-27B at 2026-10-18T07:31:57.966Z active_parameters_per_token 25.6 B — https://huggingface.co/Qwen/Qwen3.8-27B counted_at 2026-10-18T07:31:57.966Z — spec_sources.verified_at for active_params_b — when this figure was counted; a refresh that keeps it does not rewrite it confidence 0.95 — spec_sources.confidence — from the tensor headers The parameters read to produce one token. It sets the decode speed; on a dense model it equals the total. Shown to one decimal: the underlying tensor count is exact, but more digits would claim a precision the published figure does not have (SPEC §14). active	UD-Q4_K_S 15.4 GB (observed) OBSERVEDUD-Q4_K_S file size on diskMeasured or published, passed through unchanged — not computed here. quantization file size — published at the source, passed through unchanged; nothing is computed here 15.36 GB, as recorded from https://huggingface.co/unsloth/Qwen3.8-27B-GGUF/tree/main, file Qwen3.8-27B-UD-Q4_K_S.gguf quantization_file_size 15.36 GB — https://huggingface.co/unsloth/Qwen3.8-27B-GGUF/tree/main, file Qwen3.8-27B-UD-Q4_K_S.gguf verified_at 2026-10-18T07:31:57.966Z — the row's own verification timestamp quantization UD-Q4_K_S — gguf bits_per_weight 4.5682 bits — the quantizer's own naming, or measured from the file The on-disk size of the weights. It is not the memory the model occupies: add the KV cache and the runtime overhead (SPEC §8.1). Preferred over the bits-per-weight formula wherever it exists, because it includes embeddings and any higher-precision tensors. Observed, not derived: this app records the figure and its source, and does not re-measure it. on disk	18.7 GB (estimate) ESTIMATEQwen3.8-27B memory requirement at 32K contextArithmetic that also depends on a tunable assumption, listed below. required_gb = weights_gb + kv_cache_gb + overhead_gb required_gb = 15.36 + 2.15 + 1.23 = 18.73 GB weights_gb 15.358213024 GB — real file size kv_cache_gb 2.147483648 GB — kv formula overhead_gb 1.22865704192 GB — overhead fraction Real file size used instead of the params × bits formula (SPEC §8.1). 16 layers keep a full KV cache, 48 linear-attention/SSM layers keep only a constant-size state (treated as negligible). Runtime buffers, activations, and allocator slack approximated as a fraction of weights.	57 of 89 (derived) DERIVEDmachines that run Qwen3.8-27B at ≥ 15 tok/sArithmetic on observed inputs only. No tunable assumption enters it. machines = #{candidate rigs where the model loads AND clears the Plan's speed floor} 89 evaluated → 86 hold it in memory → 57 of those also reach ≥ 15 tok/s candidates_evaluated 89 machines — every complete rig the search could build from the priced catalog hold_it_in_memory 86 machines — required memory ≤ usable memory (SPEC §8.6) clear_the_speed_floor 57 machines — ≥ 15 tok/s The denominator is the machines this catalog can build today, not every machine that exists. A model with no shape or no sized quantization is counted on no machine at all, and says so rather than counting as zero.	SIZEABLE	read stale 2 d — past the 36 hour freshness window — Qwen3.8-27B: the least recent read of its data from source (shape counted Oct 18, 3:31 AM ET from the tensor headers of https://huggingface.co/Qwen/Qwen3.8-27B; quantization sizes read Oct 18, 3:31 AM ET); no upstream check after it.
SPAR3D (Stable Point-Aware 3D)Stability AI · pipeline · 3d generation	Unavailable — SPAR3D (Stable Point-Aware 3D) is a pipeline of several component models, not one network, so it has no single total parameters figure. Its size is the memory its maintainers publish for the whole pipeline. (missing: not applicable to a pipeline)	not quantized as one file	10.5 GB (observed: the maintainers’ published minimum)	53 of 89 (derived) DERIVEDmachines that run SPAR3D (Stable Point-Aware 3D)Arithmetic on observed inputs only. No tunable assumption enters it. machines = #{candidate rigs with enough memory AND a backend this pipeline supports} 89 evaluated → 53 hold it in memory → 53 of those have a backend it can run on candidates_evaluated 89 machines — every complete rig the search could build from the priced catalog hold_it_in_memory 53 machines — required memory ≤ usable memory (SPEC §8.6) supported_backend 53 machines — the CUDA / ROCm / Metal support the maintainers publish The denominator is the machines this catalog can build today, not every machine that exists. A model with no shape or no sized quantization is counted on no machine at all, and says so rather than counting as zero. A pipeline model has no token rate, so no speed floor is applied to it — only memory and backend support.	SIZEABLE	Unavailable — SPAR3D (Stable Point-Aware 3D): A pipeline model is sized by the memory and backends its maintainers publish, recorded in the seed file; nothing about it is read from Hugging Face, so it has no read date and no upstream check. (missing: not applicable to a pipeline)
Stable Fast 3D (SF3D)Stability AI · pipeline · 3d generation	Unavailable — Stable Fast 3D (SF3D) is a pipeline of several component models, not one network, so it has no single total parameters figure. Its size is the memory its maintainers publish for the whole pipeline. (missing: not applicable to a pipeline)	not quantized as one file	6.00 GB (observed: the maintainers’ published minimum)	53 of 89 (derived) DERIVEDmachines that run Stable Fast 3D (SF3D)Arithmetic on observed inputs only. No tunable assumption enters it. machines = #{candidate rigs with enough memory AND a backend this pipeline supports} 89 evaluated → 53 hold it in memory → 53 of those have a backend it can run on candidates_evaluated 89 machines — every complete rig the search could build from the priced catalog hold_it_in_memory 53 machines — required memory ≤ usable memory (SPEC §8.6) supported_backend 53 machines — the CUDA / ROCm / Metal support the maintainers publish The denominator is the machines this catalog can build today, not every machine that exists. A model with no shape or no sized quantization is counted on no machine at all, and says so rather than counting as zero. A pipeline model has no token rate, so no speed floor is applied to it — only memory and backend support.	SIZEABLE	Unavailable — Stable Fast 3D (SF3D): A pipeline model is sized by the memory and backends its maintainers publish, recorded in the seed file; nothing about it is read from Hugging Face, so it has no read date and no upstream check. (missing: not applicable to a pipeline)
In the saved Plan: Hunyuan3D 2.1 (shape + PBR paint)Tencent Hunyuan · pipeline · 3d generation	Unavailable — Hunyuan3D 2.1 (shape + PBR paint) is a pipeline of several component models, not one network, so it has no single total parameters figure. Its size is the memory its maintainers publish for the whole pipeline. (missing: not applicable to a pipeline)	not quantized as one file	29.0 GB (observed: the maintainers’ published minimum)	48 of 89 (derived) DERIVEDmachines that run Hunyuan3D 2.1 (shape + PBR paint)Arithmetic on observed inputs only. No tunable assumption enters it. machines = #{candidate rigs with enough memory AND a backend this pipeline supports} 89 evaluated → 48 hold it in memory → 48 of those have a backend it can run on candidates_evaluated 89 machines — every complete rig the search could build from the priced catalog hold_it_in_memory 48 machines — required memory ≤ usable memory (SPEC §8.6) supported_backend 48 machines — the CUDA / ROCm / Metal support the maintainers publish The denominator is the machines this catalog can build today, not every machine that exists. A model with no shape or no sized quantization is counted on no machine at all, and says so rather than counting as zero. A pipeline model has no token rate, so no speed floor is applied to it — only memory and backend support.	SIZEABLE	Unavailable — Hunyuan3D 2.1 (shape + PBR paint): A pipeline model is sized by the memory and backends its maintainers publish, recorded in the seed file; nothing about it is read from Hugging Face, so it has no read date and no upstream check. (missing: not applicable to a pipeline)
In the saved Plan: Qwen3-VL 32B InstructAlibaba Qwen · dense · vision · 256K ctx max	32.8 (derived) DERIVEDQwen3-VL 32B Instruct total parametersArithmetic on observed inputs only. No tunable assumption enters it. Σ params of language-model tensors — Cross-checked: tensor count 33.36B vs Hugging Face 33.36B (0%). 32.8 B, counted from the tensor headers of https://huggingface.co/Qwen/Qwen3-VL-32B-Instruct at 2026-10-17T12:02:07.051Z total_parameters 32.8 B — https://huggingface.co/Qwen/Qwen3-VL-32B-Instruct counted_at 2026-10-17T12:02:07.051Z — spec_sources.verified_at for total_params_b — when this figure was counted; a refresh that keeps it does not rewrite it confidence 0.95 — spec_sources.confidence — from the tensor headers The parameters of the language-model tensors only, counted from the tensor headers: a vision or audio tower and multi-token-prediction heads are excluded, so it can be below the whole-checkpoint count the formula's cross-check quotes. It sizes the weights only when no measured file size is stored for the chosen quantization; it is not what is read per token on a mixture-of-experts model. Shown to one decimal: the underlying tensor count is exact, but more digits would claim a precision the published figure does not have (SPEC §14). 32.0 (derived) DERIVEDQwen3-VL 32B Instruct active parameters per tokenArithmetic on observed inputs only. No tunable assumption enters it. active − input embedding table — Per-token memory read used for decode bandwidth (SPEC §8.2). 32.0 B, counted from the tensor headers of https://huggingface.co/Qwen/Qwen3-VL-32B-Instruct at 2026-10-17T12:02:07.051Z active_parameters_per_token 32.0 B — https://huggingface.co/Qwen/Qwen3-VL-32B-Instruct counted_at 2026-10-17T12:02:07.051Z — spec_sources.verified_at for active_params_b — when this figure was counted; a refresh that keeps it does not rewrite it confidence 0.95 — spec_sources.confidence — from the tensor headers The parameters read to produce one token. It sets the decode speed; on a dense model it equals the total. Shown to one decimal: the underlying tensor count is exact, but more digits would claim a precision the published figure does not have (SPEC §14). active	IQ4_NL 18.7 GB (observed) OBSERVEDIQ4_NL file size on diskMeasured or published, passed through unchanged — not computed here. quantization file size — published at the source, passed through unchanged; nothing is computed here 18.68 GB, as recorded from https://huggingface.co/bartowski/Qwen_Qwen3-VL-32B-Instruct-GGUF/tree/main, file Qwen_Qwen3-VL-32B-Instruct-IQ4_NL.gguf quantization_file_size 18.68 GB — https://huggingface.co/bartowski/Qwen_Qwen3-VL-32B-Instruct-GGUF/tree/main, file Qwen_Qwen3-VL-32B-Instruct-IQ4_NL.gguf verified_at 2026-10-17T12:02:07.051Z — the row's own verification timestamp quantization IQ4_NL — gguf bits_per_weight 4.5612 bits — the quantizer's own naming, or measured from the file The on-disk size of the weights. It is not the memory the model occupies: add the KV cache and the runtime overhead (SPEC §8.1). Preferred over the bits-per-weight formula wherever it exists, because it includes embeddings and any higher-precision tensors. Observed, not derived: this app records the figure and its source, and does not re-measure it. on disk	28.8 GB (estimate) ESTIMATEQwen3-VL 32B Instruct memory requirement at 32K contextArithmetic that also depends on a tunable assumption, listed below. required_gb = weights_gb + kv_cache_gb + overhead_gb required_gb = 18.68 + 8.59 + 1.49 = 28.76 GB weights_gb 18.679496 GB — real file size kv_cache_gb 8.589934592 GB — kv formula overhead_gb 1.49435968 GB — overhead fraction Real file size used instead of the params × bits formula (SPEC §8.1). 64 layers keep a full KV cache. Runtime buffers, activations, and allocator slack approximated as a fraction of weights.	48 of 89 (derived) DERIVEDmachines that run Qwen3-VL 32B Instruct at ≥ 15 tok/sArithmetic on observed inputs only. No tunable assumption enters it. machines = #{candidate rigs where the model loads AND clears the Plan's speed floor} 89 evaluated → 84 hold it in memory → 48 of those also reach ≥ 15 tok/s candidates_evaluated 89 machines — every complete rig the search could build from the priced catalog hold_it_in_memory 84 machines — required memory ≤ usable memory (SPEC §8.6) clear_the_speed_floor 48 machines — ≥ 15 tok/s The denominator is the machines this catalog can build today, not every machine that exists. A model with no shape or no sized quantization is counted on no machine at all, and says so rather than counting as zero.	SIZEABLE	read stale 3 d — past the 36 hour freshness window — Qwen3-VL 32B Instruct: the least recent read of its data from source (shape counted Oct 17, 8:02 AM ET from the tensor headers of https://huggingface.co/Qwen/Qwen3-VL-32B-Instruct; quantization sizes read Oct 17, 8:02 AM ET); never checked upstream.
InternVL3.5 38BShanghai AI Lab (OpenGVLab) · dense · vision · 40K ctx max	32.9 (derived) DERIVEDInternVL3.5 38B total parametersArithmetic on observed inputs only. No tunable assumption enters it. Σ params of language-model tensors — Cross-checked: tensor count 38.39B vs Hugging Face 38.39B (0%). 32.9 B, counted from the tensor headers of https://huggingface.co/OpenGVLab/InternVL3_5-38B at 2026-10-17T12:37:13.700Z total_parameters 32.9 B — https://huggingface.co/OpenGVLab/InternVL3_5-38B counted_at 2026-10-17T12:37:13.700Z — spec_sources.verified_at for total_params_b — when this figure was counted; a refresh that keeps it does not rewrite it confidence 0.95 — spec_sources.confidence — from the tensor headers The parameters of the language-model tensors only, counted from the tensor headers: a vision or audio tower and multi-token-prediction heads are excluded, so it can be below the whole-checkpoint count the formula's cross-check quotes. It sizes the weights only when no measured file size is stored for the chosen quantization; it is not what is read per token on a mixture-of-experts model. Shown to one decimal: the underlying tensor count is exact, but more digits would claim a precision the published figure does not have (SPEC §14). 32.1 (derived) DERIVEDInternVL3.5 38B active parameters per tokenArithmetic on observed inputs only. No tunable assumption enters it. active − input embedding table — Per-token memory read used for decode bandwidth (SPEC §8.2). 32.1 B, counted from the tensor headers of https://huggingface.co/OpenGVLab/InternVL3_5-38B at 2026-10-17T12:37:13.700Z active_parameters_per_token 32.1 B — https://huggingface.co/OpenGVLab/InternVL3_5-38B counted_at 2026-10-17T12:37:13.700Z — spec_sources.verified_at for active_params_b — when this figure was counted; a refresh that keeps it does not rewrite it confidence 0.95 — spec_sources.confidence — from the tensor headers The parameters read to produce one token. It sets the decode speed; on a dense model it equals the total. Shown to one decimal: the underlying tensor count is exact, but more digits would claim a precision the published figure does not have (SPEC §14). active	IQ4_NL 18.7 GB (observed) OBSERVEDIQ4_NL file size on diskMeasured or published, passed through unchanged — not computed here. quantization file size — published at the source, passed through unchanged; nothing is computed here 18.68 GB, as recorded from https://huggingface.co/bartowski/OpenGVLab_InternVL3_5-38B-GGUF/tree/main, file OpenGVLab_InternVL3_5-38B-IQ4_NL.gguf quantization_file_size 18.68 GB — https://huggingface.co/bartowski/OpenGVLab_InternVL3_5-38B-GGUF/tree/main, file OpenGVLab_InternVL3_5-38B-IQ4_NL.gguf verified_at 2026-10-17T12:37:13.700Z — the row's own verification timestamp quantization IQ4_NL — gguf bits_per_weight 4.5485 bits — the quantizer's own naming, or measured from the file The on-disk size of the weights. It is not the memory the model occupies: add the KV cache and the runtime overhead (SPEC §8.1). Preferred over the bits-per-weight formula wherever it exists, because it includes embeddings and any higher-precision tensors. Observed, not derived: this app records the figure and its source, and does not re-measure it. on disk	28.8 GB (estimate) ESTIMATEInternVL3.5 38B memory requirement at 32K contextArithmetic that also depends on a tunable assumption, listed below. required_gb = weights_gb + kv_cache_gb + overhead_gb required_gb = 18.68 + 8.59 + 1.49 = 28.76 GB weights_gb 18.679491584 GB — real file size kv_cache_gb 8.589934592 GB — kv formula overhead_gb 1.4943593267200002 GB — overhead fraction Real file size used instead of the params × bits formula (SPEC §8.1). 64 layers keep a full KV cache. Runtime buffers, activations, and allocator slack approximated as a fraction of weights.	48 of 89 (derived) DERIVEDmachines that run InternVL3.5 38B at ≥ 15 tok/sArithmetic on observed inputs only. No tunable assumption enters it. machines = #{candidate rigs where the model loads AND clears the Plan's speed floor} 89 evaluated → 84 hold it in memory → 48 of those also reach ≥ 15 tok/s candidates_evaluated 89 machines — every complete rig the search could build from the priced catalog hold_it_in_memory 84 machines — required memory ≤ usable memory (SPEC §8.6) clear_the_speed_floor 48 machines — ≥ 15 tok/s The denominator is the machines this catalog can build today, not every machine that exists. A model with no shape or no sized quantization is counted on no machine at all, and says so rather than counting as zero.	SIZEABLE	read stale 3 d — past the 36 hour freshness window — InternVL3.5 38B: the least recent read of its data from source (shape counted Oct 17, 8:37 AM ET from the tensor headers of https://huggingface.co/OpenGVLab/InternVL3_5-38B; quantization sizes read Oct 17, 8:37 AM ET); never checked upstream.
Step1X-3DStepFun · pipeline · 3d generation	Unavailable — Step1X-3D is a pipeline of several component models, not one network, so it has no single total parameters figure. Its size is the memory its maintainers publish for the whole pipeline. (missing: not applicable to a pipeline)	not quantized as one file	27.0 GB (observed: the maintainers’ published minimum)	35 of 89 (derived) DERIVEDmachines that run Step1X-3DArithmetic on observed inputs only. No tunable assumption enters it. machines = #{candidate rigs with enough memory AND a backend this pipeline supports} 89 evaluated → 35 hold it in memory → 35 of those have a backend it can run on candidates_evaluated 89 machines — every complete rig the search could build from the priced catalog hold_it_in_memory 35 machines — required memory ≤ usable memory (SPEC §8.6) supported_backend 35 machines — the CUDA / ROCm / Metal support the maintainers publish The denominator is the machines this catalog can build today, not every machine that exists. A model with no shape or no sized quantization is counted on no machine at all, and says so rather than counting as zero. A pipeline model has no token rate, so no speed floor is applied to it — only memory and backend support.	SIZEABLE	Unavailable — Step1X-3D: A pipeline model is sized by the memory and backends its maintainers publish, recorded in the seed file; nothing about it is read from Hugging Face, so it has no read date and no upstream check. (missing: not applicable to a pipeline)
TripoSGVAST AI Research · pipeline · 3d generation	Unavailable — TripoSG is a pipeline of several component models, not one network, so it has no single total parameters figure. Its size is the memory its maintainers publish for the whole pipeline. (missing: not applicable to a pipeline)	not quantized as one file	8.00 GB (observed: the maintainers’ published minimum)	35 of 89 (derived) DERIVEDmachines that run TripoSGArithmetic on observed inputs only. No tunable assumption enters it. machines = #{candidate rigs with enough memory AND a backend this pipeline supports} 89 evaluated → 35 hold it in memory → 35 of those have a backend it can run on candidates_evaluated 89 machines — every complete rig the search could build from the priced catalog hold_it_in_memory 35 machines — required memory ≤ usable memory (SPEC §8.6) supported_backend 35 machines — the CUDA / ROCm / Metal support the maintainers publish The denominator is the machines this catalog can build today, not every machine that exists. A model with no shape or no sized quantization is counted on no machine at all, and says so rather than counting as zero. A pipeline model has no token rate, so no speed floor is applied to it — only memory and backend support.	SIZEABLE	Unavailable — TripoSG: A pipeline model is sized by the memory and backends its maintainers publish, recorded in the seed file; nothing about it is read from Hugging Face, so it has no read date and no upstream check. (missing: not applicable to a pipeline)
GLM-4.6VZ.ai · MoE · vision · 128K ctx max	106.9 (derived) DERIVEDGLM-4.6V total parametersArithmetic on observed inputs only. No tunable assumption enters it. Σ params of language-model tensors — Cross-checked: tensor count 107.71B vs Hugging Face 107.71B (0%). 106.9 B, counted from the tensor headers of https://huggingface.co/zai-org/GLM-4.6V at 2026-10-17T12:42:49.390Z total_parameters 106.9 B — https://huggingface.co/zai-org/GLM-4.6V counted_at 2026-10-17T12:42:49.390Z — spec_sources.verified_at for total_params_b — when this figure was counted; a refresh that keeps it does not rewrite it confidence 0.95 — spec_sources.confidence — from the tensor headers The parameters of the language-model tensors only, counted from the tensor headers: a vision or audio tower and multi-token-prediction heads are excluded, so it can be below the whole-checkpoint count the formula's cross-check quotes. It sizes the weights only when no measured file size is stored for the chosen quantization; it is not what is read per token on a mixture-of-experts model. Shown to one decimal: the underlying tensor count is exact, but more digits would claim a precision the published figure does not have (SPEC §14). 12.8 (derived) DERIVEDGLM-4.6V active parameters per tokenArithmetic on observed inputs only. No tunable assumption enters it. active − input embedding table — Per-token memory read used for decode bandwidth (SPEC §8.2). 12.8 B, counted from the tensor headers of https://huggingface.co/zai-org/GLM-4.6V at 2026-10-17T12:42:49.390Z active_parameters_per_token 12.8 B — https://huggingface.co/zai-org/GLM-4.6V counted_at 2026-10-17T12:42:49.390Z — spec_sources.verified_at for active_params_b — when this figure was counted; a refresh that keeps it does not rewrite it confidence 0.95 — spec_sources.confidence — from the tensor headers The parameters read to produce one token. It sets the decode speed; on a dense model it equals the total. Shown to one decimal: the underlying tensor count is exact, but more digits would claim a precision the published figure does not have (SPEC §14). active	IQ4_NL 60.6 GB (observed) OBSERVEDIQ4_NL file size on diskMeasured or published, passed through unchanged — not computed here. quantization file size — published at the source, passed through unchanged; nothing is computed here 60.58 GB, as recorded from https://huggingface.co/unsloth/GLM-4.6V-GGUF/tree/main, file 2 files quantization_file_size 60.58 GB — https://huggingface.co/unsloth/GLM-4.6V-GGUF/tree/main, file 2 files verified_at 2026-10-17T12:42:49.390Z — the row's own verification timestamp quantization IQ4_NL — gguf bits_per_weight 4.5354 bits — the quantizer's own naming, or measured from the file The on-disk size of the weights. It is not the memory the model occupies: add the KV cache and the runtime overhead (SPEC §8.1). Preferred over the bits-per-weight formula wherever it exists, because it includes embeddings and any higher-precision tensors. Observed, not derived: this app records the figure and its source, and does not re-measure it. on disk	71.6 GB (estimate) ESTIMATEGLM-4.6V memory requirement at 32K contextArithmetic that also depends on a tunable assumption, listed below. required_gb = weights_gb + kv_cache_gb + overhead_gb required_gb = 60.58 + 6.17 + 4.85 = 71.6 GB weights_gb 60.577246112 GB — real file size kv_cache_gb 6.174015488 GB — kv formula overhead_gb 4.8461796889599995 GB — overhead fraction Real file size used instead of the params × bits formula (SPEC §8.1). 46 layers keep a full KV cache. Runtime buffers, activations, and allocator slack approximated as a fraction of weights.	28 of 89 (derived) DERIVEDmachines that run GLM-4.6V at ≥ 15 tok/sArithmetic on observed inputs only. No tunable assumption enters it. machines = #{candidate rigs where the model loads AND clears the Plan's speed floor} 89 evaluated → 74 hold it in memory → 28 of those also reach ≥ 15 tok/s candidates_evaluated 89 machines — every complete rig the search could build from the priced catalog hold_it_in_memory 74 machines — required memory ≤ usable memory (SPEC §8.6) clear_the_speed_floor 28 machines — ≥ 15 tok/s The denominator is the machines this catalog can build today, not every machine that exists. A model with no shape or no sized quantization is counted on no machine at all, and says so rather than counting as zero.	SIZEABLE	read stale 3 d — past the 36 hour freshness window — GLM-4.6V: the least recent read of its data from source (shape counted Oct 17, 8:42 AM ET from the tensor headers of https://huggingface.co/zai-org/GLM-4.6V; quantization sizes read Oct 17, 8:42 AM ET); never checked upstream.
GLM-5.3-FlashZ.ai (zai-org) · MoE · coding, vision · 1M ctx max	313.3 (derived) DERIVEDGLM-5.3-Flash total parametersArithmetic on observed inputs only. No tunable assumption enters it. Σ params of language-model tensors — Cross-checked: tensor count 321.32B vs Hugging Face 321.32B (0%). 313.3 B, counted from the tensor headers of https://huggingface.co/zai-org/GLM-5.3-Flash at 2026-10-18T07:27:57.539Z total_parameters 313.3 B — https://huggingface.co/zai-org/GLM-5.3-Flash counted_at 2026-10-18T07:27:57.539Z — spec_sources.verified_at for total_params_b — when this figure was counted; a refresh that keeps it does not rewrite it confidence 0.95 — spec_sources.confidence — from the tensor headers The parameters of the language-model tensors only, counted from the tensor headers: a vision or audio tower and multi-token-prediction heads are excluded, so it can be below the whole-checkpoint count the formula's cross-check quotes. It sizes the weights only when no measured file size is stored for the chosen quantization; it is not what is read per token on a mixture-of-experts model. Shown to one decimal: the underlying tensor count is exact, but more digits would claim a precision the published figure does not have (SPEC §14). 16.7 (derived) DERIVEDGLM-5.3-Flash active parameters per tokenArithmetic on observed inputs only. No tunable assumption enters it. active − input embedding table — Per-token memory read used for decode bandwidth (SPEC §8.2). 16.7 B, counted from the tensor headers of https://huggingface.co/zai-org/GLM-5.3-Flash at 2026-10-18T07:27:57.539Z active_parameters_per_token 16.7 B — https://huggingface.co/zai-org/GLM-5.3-Flash counted_at 2026-10-18T07:27:57.539Z — spec_sources.verified_at for active_params_b — when this figure was counted; a refresh that keeps it does not rewrite it confidence 0.95 — spec_sources.confidence — from the tensor headers The parameters read to produce one token. It sets the decode speed; on a dense model it equals the total. Shown to one decimal: the underlying tensor count is exact, but more digits would claim a precision the published figure does not have (SPEC §14). active	UD-IQ4_XS 157 GB (observed) OBSERVEDUD-IQ4_XS file size on diskMeasured or published, passed through unchanged — not computed here. quantization file size — published at the source, passed through unchanged; nothing is computed here 156.82 GB, as recorded from https://huggingface.co/unsloth/GLM-5.3-Flash-GGUF/tree/main, file 5 files quantization_file_size 156.82 GB — https://huggingface.co/unsloth/GLM-5.3-Flash-GGUF/tree/main, file 5 files verified_at 2026-10-18T07:27:57.539Z — the row's own verification timestamp quantization UD-IQ4_XS — gguf bits_per_weight 4.0041 bits — the quantizer's own naming, or measured from the file The on-disk size of the weights. It is not the memory the model occupies: add the KV cache and the runtime overhead (SPEC §8.1). Preferred over the bits-per-weight formula wherever it exists, because it includes embeddings and any higher-precision tensors. Observed, not derived: this app records the figure and its source, and does not re-measure it. on disk	170 GB (estimate) ESTIMATEGLM-5.3-Flash memory requirement at 32K contextArithmetic that also depends on a tunable assumption, listed below. required_gb = weights_gb + kv_cache_gb + overhead_gb required_gb = 156.82 + 0.37 + 12.55 = 169.74 GB weights_gb 156.8221112 GB — real file size kv_cache_gb 0.369098752 GB — kv formula overhead_gb 12.545768896 GB — overhead fraction Real file size used instead of the params × bits formula (SPEC §8.1). 11 layers keep a full KV cache, 34 linear-attention/SSM layers keep only a constant-size state (treated as negligible). Runtime buffers, activations, and allocator slack approximated as a fraction of weights.	22 of 89 (derived) DERIVEDmachines that run GLM-5.3-Flash at ≥ 15 tok/sArithmetic on observed inputs only. No tunable assumption enters it. machines = #{candidate rigs where the model loads AND clears the Plan's speed floor} 89 evaluated → 54 hold it in memory → 22 of those also reach ≥ 15 tok/s candidates_evaluated 89 machines — every complete rig the search could build from the priced catalog hold_it_in_memory 54 machines — required memory ≤ usable memory (SPEC §8.6) clear_the_speed_floor 22 machines — ≥ 15 tok/s The denominator is the machines this catalog can build today, not every machine that exists. A model with no shape or no sized quantization is counted on no machine at all, and says so rather than counting as zero.	SIZEABLE	read stale 2 d — past the 36 hour freshness window — GLM-5.3-Flash: the least recent read of its data from source (shape counted Oct 18, 3:27 AM ET from the tensor headers of https://huggingface.co/zai-org/GLM-5.3-Flash; quantization sizes read Oct 18, 3:27 AM ET); no upstream check after it.
DeepSeek-V4-Flash-0731DeepSeek · MoE · reasoning, coding · 1M ctx max	284.3 (derived) DERIVEDDeepSeek-V4-Flash-0731 total parametersArithmetic on observed inputs only. No tunable assumption enters it. Σ params of language-model tensors — Cross-checked: tensor count 304.18B vs Hugging Face 304.18B (-7.65e-4%). 284.3 B, counted from the tensor headers of https://huggingface.co/deepseek-ai/DeepSeek-V4-Flash-0731 at 2026-10-17T09:35:09.394Z total_parameters 284.3 B — https://huggingface.co/deepseek-ai/DeepSeek-V4-Flash-0731 counted_at 2026-10-17T09:35:09.394Z — spec_sources.verified_at for total_params_b — when this figure was counted; a refresh that keeps it does not rewrite it confidence 0.95 — spec_sources.confidence — from the tensor headers The parameters of the language-model tensors only, counted from the tensor headers: a vision or audio tower and multi-token-prediction heads are excluded, so it can be below the whole-checkpoint count the formula's cross-check quotes. It sizes the weights only when no measured file size is stored for the chosen quantization; it is not what is read per token on a mixture-of-experts model. Shown to one decimal: the underlying tensor count is exact, but more digits would claim a precision the published figure does not have (SPEC §14). 13.3 (derived) DERIVEDDeepSeek-V4-Flash-0731 active parameters per tokenArithmetic on observed inputs only. No tunable assumption enters it. active − input embedding table — Per-token memory read used for decode bandwidth (SPEC §8.2). 13.3 B, counted from the tensor headers of https://huggingface.co/deepseek-ai/DeepSeek-V4-Flash-0731 at 2026-10-17T09:35:09.394Z active_parameters_per_token 13.3 B — https://huggingface.co/deepseek-ai/DeepSeek-V4-Flash-0731 counted_at 2026-10-17T09:35:09.394Z — spec_sources.verified_at for active_params_b — when this figure was counted; a refresh that keeps it does not rewrite it confidence 0.95 — spec_sources.confidence — from the tensor headers The parameters read to produce one token. It sets the decode speed; on a dense model it equals the total. Shown to one decimal: the underlying tensor count is exact, but more digits would claim a precision the published figure does not have (SPEC §14). active	UD-Q8_K_XL 162 GB (observed) OBSERVEDUD-Q8_K_XL file size on diskMeasured or published, passed through unchanged — not computed here. quantization file size — published at the source, passed through unchanged; nothing is computed here 161.87 GB, as recorded from https://huggingface.co/unsloth/DeepSeek-V4-Flash-0731-GGUF/tree/main, file 5 files quantization_file_size 161.87 GB — https://huggingface.co/unsloth/DeepSeek-V4-Flash-0731-GGUF/tree/main, file 5 files verified_at 2026-10-17T09:35:09.394Z — the row's own verification timestamp quantization UD-Q8_K_XL — gguf bits_per_weight 4.5544 bits — the quantizer's own naming, or measured from the file The on-disk size of the weights. It is not the memory the model occupies: add the KV cache and the runtime overhead (SPEC §8.1). Preferred over the bits-per-weight formula wherever it exists, because it includes embeddings and any higher-precision tensors. Observed, not derived: this app records the figure and its source, and does not re-measure it. on disk	176 GB (estimate) ESTIMATEDeepSeek-V4-Flash-0731 memory requirement at 32K contextArithmetic that also depends on a tunable assumption, listed below. required_gb = weights_gb + kv_cache_gb + overhead_gb required_gb = 161.87 + 1.44 + 12.95 = 176.26 GB weights_gb 161.86961552 GB — real file size kv_cache_gb 1.442840576 GB — kv formula overhead_gb 12.9495692416 GB — overhead fraction Real file size used instead of the params × bits formula (SPEC §8.1). 43 layers keep a full KV cache. KV width inferred from a single latent 'wkv' projection (unrecognized attention family); any KV compression the runtime applies is not modeled, so the cache is likely overstated. Runtime buffers, activations, and allocator slack approximated as a fraction of weights.	18 of 89 (derived) DERIVEDmachines that run DeepSeek-V4-Flash-0731 at ≥ 15 tok/sArithmetic on observed inputs only. No tunable assumption enters it. machines = #{candidate rigs where the model loads AND clears the Plan's speed floor} 89 evaluated → 53 hold it in memory → 18 of those also reach ≥ 15 tok/s candidates_evaluated 89 machines — every complete rig the search could build from the priced catalog hold_it_in_memory 53 machines — required memory ≤ usable memory (SPEC §8.6) clear_the_speed_floor 18 machines — ≥ 15 tok/s The denominator is the machines this catalog can build today, not every machine that exists. A model with no shape or no sized quantization is counted on no machine at all, and says so rather than counting as zero.	SIZEABLE	read stale 3 d — past the 36 hour freshness window — DeepSeek-V4-Flash-0731: the least recent read of its data from source (shape counted Oct 17, 5:35 AM ET from the tensor headers of https://huggingface.co/deepseek-ai/DeepSeek-V4-Flash-0731; quantization sizes read Oct 17, 5:35 AM ET); never checked upstream.
In the saved Plan: Qwen3-ASR 1.7BAlibaba Qwen · dense · speech to text	2.0 (derived) DERIVEDQwen3-ASR 1.7B total parametersArithmetic on observed inputs only. No tunable assumption enters it. Σ params of language-model tensors — Cross-checked: tensor count 2.35B vs Hugging Face 2.35B (0%). 2.0 B, counted from the tensor headers of https://huggingface.co/Qwen/Qwen3-ASR-1.7B at 2026-10-17T11:50:02.565Z total_parameters 2.0 B — https://huggingface.co/Qwen/Qwen3-ASR-1.7B counted_at 2026-10-17T11:50:02.565Z — spec_sources.verified_at for total_params_b — when this figure was counted; a refresh that keeps it does not rewrite it confidence 0.95 — spec_sources.confidence — from the tensor headers The parameters of the language-model tensors only, counted from the tensor headers: a vision or audio tower and multi-token-prediction heads are excluded, so it can be below the whole-checkpoint count the formula's cross-check quotes. It sizes the weights only when no measured file size is stored for the chosen quantization; it is not what is read per token on a mixture-of-experts model. Shown to one decimal: the underlying tensor count is exact, but more digits would claim a precision the published figure does not have (SPEC §14). 1.7 (derived) DERIVEDQwen3-ASR 1.7B active parameters per tokenArithmetic on observed inputs only. No tunable assumption enters it. active − input embedding table — Per-token memory read used for decode bandwidth (SPEC §8.2). 1.7 B, counted from the tensor headers of https://huggingface.co/Qwen/Qwen3-ASR-1.7B at 2026-10-17T11:50:02.565Z active_parameters_per_token 1.7 B — https://huggingface.co/Qwen/Qwen3-ASR-1.7B counted_at 2026-10-17T11:50:02.565Z — spec_sources.verified_at for active_params_b — when this figure was counted; a refresh that keeps it does not rewrite it confidence 0.95 — spec_sources.confidence — from the tensor headers The parameters read to produce one token. It sets the decode speed; on a dense model it equals the total. Shown to one decimal: the underlying tensor count is exact, but more digits would claim a precision the published figure does not have (SPEC §14). active	6bit 2.03 GB (observed) OBSERVED6bit file size on diskMeasured or published, passed through unchanged — not computed here. quantization file size — published at the source, passed through unchanged; nothing is computed here 2.03 GB, as recorded from https://huggingface.co/mlx-community/Mega-ASR-6bit/tree/main, file model.safetensors quantization_file_size 2.03 GB — https://huggingface.co/mlx-community/Mega-ASR-6bit/tree/main, file model.safetensors verified_at 2026-10-17T11:50:02.565Z — the row's own verification timestamp quantization 6bit — mlx bits_per_weight 8.0057 bits — the quantizer's own naming, or measured from the file The on-disk size of the weights. It is not the memory the model occupies: add the KV cache and the runtime overhead (SPEC §8.1). Preferred over the bits-per-weight formula wherever it exists, because it includes embeddings and any higher-precision tensors. Observed, not derived: this app records the figure and its source, and does not re-measure it. on disk	5.95 GB (estimate) ESTIMATEQwen3-ASR 1.7B memory requirement at 32K contextArithmetic that also depends on a tunable assumption, listed below. required_gb = weights_gb + kv_cache_gb + overhead_gb required_gb = 2.03 + 3.76 + 0.16 = 5.95 GB weights_gb 2.033194557 GB — real file size kv_cache_gb 3.758096384 GB — kv formula overhead_gb 0.16265556456 GB — overhead fraction Real file size used instead of the params × bits formula (SPEC §8.1). 28 layers keep a full KV cache. Runtime buffers, activations, and allocator slack approximated as a fraction of weights.	18 of 89 (derived) DERIVEDmachines that run Qwen3-ASR 1.7B at ≥ 15 tok/sArithmetic on observed inputs only. No tunable assumption enters it. machines = #{candidate rigs where the model loads AND clears the Plan's speed floor} 89 evaluated → 18 hold it in memory → 18 of those also reach ≥ 15 tok/s candidates_evaluated 89 machines — every complete rig the search could build from the priced catalog hold_it_in_memory 18 machines — required memory ≤ usable memory (SPEC §8.6) clear_the_speed_floor 18 machines — ≥ 15 tok/s The denominator is the machines this catalog can build today, not every machine that exists. A model with no shape or no sized quantization is counted on no machine at all, and says so rather than counting as zero.	SIZEABLE	read stale 3 d — past the 36 hour freshness window — Qwen3-ASR 1.7B: the least recent read of its data from source (shape counted Oct 17, 7:50 AM ET from the tensor headers of https://huggingface.co/Qwen/Qwen3-ASR-1.7B; quantization sizes read Oct 17, 7:50 AM ET); never checked upstream.
Unlimited-OCRBaidu · MoE · vision · 32K ctx max	3.0 (derived) DERIVEDUnlimited-OCR total parametersArithmetic on observed inputs only. No tunable assumption enters it. Σ params of language-model tensors — Cross-checked: tensor count 3.34B vs Hugging Face 3.34B (0%). 3.0 B, counted from the tensor headers of https://huggingface.co/baidu/Unlimited-OCR at 2026-10-17T13:09:50.550Z total_parameters 3.0 B — https://huggingface.co/baidu/Unlimited-OCR counted_at 2026-10-17T13:09:50.550Z — spec_sources.verified_at for total_params_b — when this figure was counted; a refresh that keeps it does not rewrite it confidence 0.95 — spec_sources.confidence — from the tensor headers The parameters of the language-model tensors only, counted from the tensor headers: a vision or audio tower and multi-token-prediction heads are excluded, so it can be below the whole-checkpoint count the formula's cross-check quotes. It sizes the weights only when no measured file size is stored for the chosen quantization; it is not what is read per token on a mixture-of-experts model. Shown to one decimal: the underlying tensor count is exact, but more digits would claim a precision the published figure does not have (SPEC §14). 0.7 (derived) DERIVEDUnlimited-OCR active parameters per tokenArithmetic on observed inputs only. No tunable assumption enters it. active − input embedding table — Per-token memory read used for decode bandwidth (SPEC §8.2). 0.7 B, counted from the tensor headers of https://huggingface.co/baidu/Unlimited-OCR at 2026-10-17T13:09:50.550Z active_parameters_per_token 0.7 B — https://huggingface.co/baidu/Unlimited-OCR counted_at 2026-10-17T13:09:50.550Z — spec_sources.verified_at for active_params_b — when this figure was counted; a refresh that keeps it does not rewrite it confidence 0.95 — spec_sources.confidence — from the tensor headers The parameters read to produce one token. It sets the decode speed; on a dense model it equals the total. Shown to one decimal: the underlying tensor count is exact, but more digits would claim a precision the published figure does not have (SPEC §14). active	4bit 2.45 GB (observed) OBSERVED4bit file size on diskMeasured or published, passed through unchanged — not computed here. quantization file size — published at the source, passed through unchanged; nothing is computed here 2.45 GB, as recorded from https://huggingface.co/mlx-community/Unlimited-OCR-4bit/tree/main, file model.safetensors quantization_file_size 2.45 GB — https://huggingface.co/mlx-community/Unlimited-OCR-4bit/tree/main, file model.safetensors verified_at 2026-10-17T13:09:50.550Z — the row's own verification timestamp quantization 4bit — mlx bits_per_weight 6.4656 bits — the quantizer's own naming, or measured from the file The on-disk size of the weights. It is not the memory the model occupies: add the KV cache and the runtime overhead (SPEC §8.1). Preferred over the bits-per-weight formula wherever it exists, because it includes embeddings and any higher-precision tensors. Observed, not derived: this app records the figure and its source, and does not re-measure it. on disk	4.66 GB (estimate) ESTIMATEUnlimited-OCR memory requirement at 32K contextArithmetic that also depends on a tunable assumption, listed below. required_gb = weights_gb + kv_cache_gb + overhead_gb required_gb = 2.45 + 2.01 + 0.2 = 4.66 GB weights_gb 2.451215084 GB — real file size kv_cache_gb 2.01326592 GB — kv formula overhead_gb 0.19609720672000003 GB — overhead fraction Real file size used instead of the params × bits formula (SPEC §8.1). 12 layers keep a full KV cache. Runtime buffers, activations, and allocator slack approximated as a fraction of weights.	18 of 89 (derived) DERIVEDmachines that run Unlimited-OCR at ≥ 15 tok/sArithmetic on observed inputs only. No tunable assumption enters it. machines = #{candidate rigs where the model loads AND clears the Plan's speed floor} 89 evaluated → 18 hold it in memory → 18 of those also reach ≥ 15 tok/s candidates_evaluated 89 machines — every complete rig the search could build from the priced catalog hold_it_in_memory 18 machines — required memory ≤ usable memory (SPEC §8.6) clear_the_speed_floor 18 machines — ≥ 15 tok/s The denominator is the machines this catalog can build today, not every machine that exists. A model with no shape or no sized quantization is counted on no machine at all, and says so rather than counting as zero.	SIZEABLE	read stale 3 d — past the 36 hour freshness window — Unlimited-OCR: the least recent read of its data from source (shape counted Oct 17, 9:09 AM ET from the tensor headers of https://huggingface.co/baidu/Unlimited-OCR; quantization sizes read Oct 17, 9:09 AM ET); never checked upstream.
Qwen3-VL 235B-A22B Instruct — shown in full belowAlibaba Qwen · MoE · vision · 256K ctx max	235.1 (derived) DERIVEDQwen3-VL 235B-A22B Instruct total parametersArithmetic on observed inputs only. No tunable assumption enters it. Σ params of language-model tensors — Cross-checked: tensor count 235.67B vs Hugging Face 235.67B (0%). 235.1 B, counted from the tensor headers of https://huggingface.co/Qwen/Qwen3-VL-235B-A22B-Instruct at 2026-10-17T12:11:06.530Z total_parameters 235.1 B — https://huggingface.co/Qwen/Qwen3-VL-235B-A22B-Instruct counted_at 2026-10-17T12:11:06.530Z — spec_sources.verified_at for total_params_b — when this figure was counted; a refresh that keeps it does not rewrite it confidence 0.95 — spec_sources.confidence — from the tensor headers The parameters of the language-model tensors only, counted from the tensor headers: a vision or audio tower and multi-token-prediction heads are excluded, so it can be below the whole-checkpoint count the formula's cross-check quotes. It sizes the weights only when no measured file size is stored for the chosen quantization; it is not what is read per token on a mixture-of-experts model. Shown to one decimal: the underlying tensor count is exact, but more digits would claim a precision the published figure does not have (SPEC §14). 21.6 (derived) DERIVEDQwen3-VL 235B-A22B Instruct active parameters per tokenArithmetic on observed inputs only. No tunable assumption enters it. active − input embedding table — Per-token memory read used for decode bandwidth (SPEC §8.2). 21.6 B, counted from the tensor headers of https://huggingface.co/Qwen/Qwen3-VL-235B-A22B-Instruct at 2026-10-17T12:11:06.530Z active_parameters_per_token 21.6 B — https://huggingface.co/Qwen/Qwen3-VL-235B-A22B-Instruct counted_at 2026-10-17T12:11:06.530Z — spec_sources.verified_at for active_params_b — when this figure was counted; a refresh that keeps it does not rewrite it confidence 0.95 — spec_sources.confidence — from the tensor headers The parameters read to produce one token. It sets the decode speed; on a dense model it equals the total. Shown to one decimal: the underlying tensor count is exact, but more digits would claim a precision the published figure does not have (SPEC §14). active	IQ4_NL 133 GB (observed) OBSERVEDIQ4_NL file size on diskMeasured or published, passed through unchanged — not computed here. quantization file size — published at the source, passed through unchanged; nothing is computed here 132.60 GB, as recorded from https://huggingface.co/unsloth/Qwen3-VL-235B-A22B-Instruct-1M-GGUF/tree/main, file 3 files quantization_file_size 132.60 GB — https://huggingface.co/unsloth/Qwen3-VL-235B-A22B-Instruct-1M-GGUF/tree/main, file 3 files verified_at 2026-10-17T12:11:06.530Z — the row's own verification timestamp quantization IQ4_NL — gguf bits_per_weight 4.5124 bits — the quantizer's own naming, or measured from the file The on-disk size of the weights. It is not the memory the model occupies: add the KV cache and the runtime overhead (SPEC §8.1). Preferred over the bits-per-weight formula wherever it exists, because it includes embeddings and any higher-precision tensors. Observed, not derived: this app records the figure and its source, and does not re-measure it. on disk	150 GB (estimate) ESTIMATEQwen3-VL 235B-A22B Instruct memory requirement at 32K contextArithmetic that also depends on a tunable assumption, listed below. required_gb = weights_gb + kv_cache_gb + overhead_gb required_gb = 132.6 + 6.31 + 10.61 = 149.52 GB weights_gb 132.60341664 GB — real file size kv_cache_gb 6.308233216 GB — kv formula overhead_gb 10.608273331200001 GB — overhead fraction Real file size used instead of the params × bits formula (SPEC §8.1). 94 layers keep a full KV cache. Runtime buffers, activations, and allocator slack approximated as a fraction of weights.	14 of 89 (derived) DERIVEDmachines that run Qwen3-VL 235B-A22B Instruct at ≥ 15 tok/sArithmetic on observed inputs only. No tunable assumption enters it. machines = #{candidate rigs where the model loads AND clears the Plan's speed floor} 89 evaluated → 54 hold it in memory → 14 of those also reach ≥ 15 tok/s candidates_evaluated 89 machines — every complete rig the search could build from the priced catalog hold_it_in_memory 54 machines — required memory ≤ usable memory (SPEC §8.6) clear_the_speed_floor 14 machines — ≥ 15 tok/s The denominator is the machines this catalog can build today, not every machine that exists. A model with no shape or no sized quantization is counted on no machine at all, and says so rather than counting as zero.	SIZEABLE	read stale 3 d — past the 36 hour freshness window — Qwen3-VL 235B-A22B Instruct: the least recent read of its data from source (shape counted Oct 17, 8:11 AM ET from the tensor headers of https://huggingface.co/Qwen/Qwen3-VL-235B-A22B-Instruct; quantization sizes read Oct 17, 8:11 AM ET); never checked upstream.
Qwen3.8-Flash-NextAlibaba Qwen · MoE · coding, vision · 256K ctx max	176.9 (derived) DERIVEDQwen3.8-Flash-Next total parametersArithmetic on observed inputs only. No tunable assumption enters it. Σ params of language-model tensors — Cross-checked: tensor count 180B vs Hugging Face 180B (0%). 176.9 B, counted from the tensor headers of https://huggingface.co/Qwen/Qwen3.8-Flash-Next at 2026-10-18T07:33:12.113Z total_parameters 176.9 B — https://huggingface.co/Qwen/Qwen3.8-Flash-Next counted_at 2026-10-18T07:33:12.113Z — spec_sources.verified_at for total_params_b — when this figure was counted; a refresh that keeps it does not rewrite it confidence 0.95 — spec_sources.confidence — from the tensor headers The parameters of the language-model tensors only, counted from the tensor headers: a vision or audio tower and multi-token-prediction heads are excluded, so it can be below the whole-checkpoint count the formula's cross-check quotes. It sizes the weights only when no measured file size is stored for the chosen quantization; it is not what is read per token on a mixture-of-experts model. Shown to one decimal: the underlying tensor count is exact, but more digits would claim a precision the published figure does not have (SPEC §14). 57.9 (derived) DERIVEDQwen3.8-Flash-Next active parameters per tokenArithmetic on observed inputs only. No tunable assumption enters it. active − input embedding table — Per-token memory read used for decode bandwidth (SPEC §8.2). 57.9 B, counted from the tensor headers of https://huggingface.co/Qwen/Qwen3.8-Flash-Next at 2026-10-18T07:33:12.113Z active_parameters_per_token 57.9 B — https://huggingface.co/Qwen/Qwen3.8-Flash-Next counted_at 2026-10-18T07:33:12.113Z — spec_sources.verified_at for active_params_b — when this figure was counted; a refresh that keeps it does not rewrite it confidence 0.95 — spec_sources.confidence — from the tensor headers The parameters read to produce one token. It sets the decode speed; on a dense model it equals the total. Shown to one decimal: the underlying tensor count is exact, but more digits would claim a precision the published figure does not have (SPEC §14). active	IQ4_NL 100 GB (observed) OBSERVEDIQ4_NL file size on diskMeasured or published, passed through unchanged — not computed here. quantization file size — published at the source, passed through unchanged; nothing is computed here 100.29 GB, as recorded from https://huggingface.co/bartowski/Qwen3.8-Flash-Next-GGUF/tree/main, file 3 files quantization_file_size 100.29 GB — https://huggingface.co/bartowski/Qwen3.8-Flash-Next-GGUF/tree/main, file 3 files verified_at 2026-10-18T07:33:12.113Z — the row's own verification timestamp quantization IQ4_NL — gguf bits_per_weight 4.5345 bits — the quantizer's own naming, or measured from the file The on-disk size of the weights. It is not the memory the model occupies: add the KV cache and the runtime overhead (SPEC §8.1). Preferred over the bits-per-weight formula wherever it exists, because it includes embeddings and any higher-precision tensors. Observed, not derived: this app records the figure and its source, and does not re-measure it. on disk	109 GB (estimate) ESTIMATEQwen3.8-Flash-Next memory requirement at 32K contextArithmetic that also depends on a tunable assumption, listed below. required_gb = weights_gb + kv_cache_gb + overhead_gb required_gb = 100.29 + 0.81 + 8.02 = 109.12 GB weights_gb 100.293527232 GB — real file size kv_cache_gb 0.805306368 GB — kv formula overhead_gb 8.02348217856 GB — overhead fraction Real file size used instead of the params × bits formula (SPEC §8.1). 12 layers keep a full KV cache, 36 linear-attention/SSM layers keep only a constant-size state (treated as negligible). Runtime buffers, activations, and allocator slack approximated as a fraction of weights.	14 of 89 (derived) DERIVEDmachines that run Qwen3.8-Flash-Next at ≥ 15 tok/sArithmetic on observed inputs only. No tunable assumption enters it. machines = #{candidate rigs where the model loads AND clears the Plan's speed floor} 89 evaluated → 69 hold it in memory → 14 of those also reach ≥ 15 tok/s candidates_evaluated 89 machines — every complete rig the search could build from the priced catalog hold_it_in_memory 69 machines — required memory ≤ usable memory (SPEC §8.6) clear_the_speed_floor 14 machines — ≥ 15 tok/s The denominator is the machines this catalog can build today, not every machine that exists. A model with no shape or no sized quantization is counted on no machine at all, and says so rather than counting as zero.	SIZEABLE	read stale 2 d — past the 36 hour freshness window — Qwen3.8-Flash-Next: the least recent read of its data from source (shape counted Oct 18, 3:33 AM ET from the tensor headers of https://huggingface.co/Qwen/Qwen3.8-Flash-Next; quantization sizes read Oct 18, 3:33 AM ET); no upstream check after it.
Mistral-Medium-3.5-128BMistral AI · dense · coding, reasoning, vision · 256K ctx max	125.0 (derived) DERIVEDMistral-Medium-3.5-128B total parametersArithmetic on observed inputs only. No tunable assumption enters it. Σ params of language-model tensors — Cross-checked: tensor count 127.7B vs Hugging Face 127.7B (4.82e-7%). 125.0 B, counted from the tensor headers of https://huggingface.co/mistralai/Mistral-Medium-3.5-128B at 2026-10-17T11:23:22.215Z total_parameters 125.0 B — https://huggingface.co/mistralai/Mistral-Medium-3.5-128B counted_at 2026-10-17T11:23:22.215Z — spec_sources.verified_at for total_params_b — when this figure was counted; a refresh that keeps it does not rewrite it confidence 0.95 — spec_sources.confidence — from the tensor headers The parameters of the language-model tensors only, counted from the tensor headers: a vision or audio tower and multi-token-prediction heads are excluded, so it can be below the whole-checkpoint count the formula's cross-check quotes. It sizes the weights only when no measured file size is stored for the chosen quantization; it is not what is read per token on a mixture-of-experts model. Shown to one decimal: the underlying tensor count is exact, but more digits would claim a precision the published figure does not have (SPEC §14). 123.4 (derived) DERIVEDMistral-Medium-3.5-128B active parameters per tokenArithmetic on observed inputs only. No tunable assumption enters it. active − input embedding table — Per-token memory read used for decode bandwidth (SPEC §8.2). 123.4 B, counted from the tensor headers of https://huggingface.co/mistralai/Mistral-Medium-3.5-128B at 2026-10-17T11:23:22.215Z active_parameters_per_token 123.4 B — https://huggingface.co/mistralai/Mistral-Medium-3.5-128B counted_at 2026-10-17T11:23:22.215Z — spec_sources.verified_at for active_params_b — when this figure was counted; a refresh that keeps it does not rewrite it confidence 0.95 — spec_sources.confidence — from the tensor headers The parameters read to produce one token. It sets the decode speed; on a dense model it equals the total. Shown to one decimal: the underlying tensor count is exact, but more digits would claim a precision the published figure does not have (SPEC §14). active	IQ4_NL 70.9 GB (observed) OBSERVEDIQ4_NL file size on diskMeasured or published, passed through unchanged — not computed here. quantization file size — published at the source, passed through unchanged; nothing is computed here 70.90 GB, as recorded from https://huggingface.co/unsloth/Mistral-Medium-3.5-128B-GGUF/tree/main, file 3 files quantization_file_size 70.90 GB — https://huggingface.co/unsloth/Mistral-Medium-3.5-128B-GGUF/tree/main, file 3 files verified_at 2026-10-17T11:23:22.215Z — the row's own verification timestamp quantization IQ4_NL — gguf bits_per_weight 4.5364 bits — the quantizer's own naming, or measured from the file The on-disk size of the weights. It is not the memory the model occupies: add the KV cache and the runtime overhead (SPEC §8.1). Preferred over the bits-per-weight formula wherever it exists, because it includes embeddings and any higher-precision tensors. Observed, not derived: this app records the figure and its source, and does not re-measure it. on disk	88.4 GB (estimate) ESTIMATEMistral-Medium-3.5-128B memory requirement at 32K contextArithmetic that also depends on a tunable assumption, listed below. required_gb = weights_gb + kv_cache_gb + overhead_gb required_gb = 70.9 + 11.81 + 5.67 = 88.38 GB weights_gb 70.896166304 GB — real file size kv_cache_gb 11.811160064 GB — kv formula overhead_gb 5.671693304320001 GB — overhead fraction Real file size used instead of the params × bits formula (SPEC §8.1). 88 layers keep a full KV cache. Runtime buffers, activations, and allocator slack approximated as a fraction of weights.	11 of 89 (derived) DERIVEDmachines that run Mistral-Medium-3.5-128B at ≥ 15 tok/sArithmetic on observed inputs only. No tunable assumption enters it. machines = #{candidate rigs where the model loads AND clears the Plan's speed floor} 89 evaluated → 70 hold it in memory → 11 of those also reach ≥ 15 tok/s candidates_evaluated 89 machines — every complete rig the search could build from the priced catalog hold_it_in_memory 70 machines — required memory ≤ usable memory (SPEC §8.6) clear_the_speed_floor 11 machines — ≥ 15 tok/s The denominator is the machines this catalog can build today, not every machine that exists. A model with no shape or no sized quantization is counted on no machine at all, and says so rather than counting as zero.	SIZEABLE	read stale 3 d — past the 36 hour freshness window — Mistral-Medium-3.5-128B: the least recent read of its data from source (shape counted Oct 17, 7:23 AM ET from the tensor headers of https://huggingface.co/mistralai/Mistral-Medium-3.5-128B; quantization sizes read Oct 17, 7:23 AM ET); never checked upstream.
Qwen3-Coder-NextAlibaba Qwen · MoE · coding · 256K ctx max	79.7 (derived) DERIVEDQwen3-Coder-Next total parametersArithmetic on observed inputs only. No tunable assumption enters it. Σ params of language-model tensors — Cross-checked: tensor count 79.67B vs Hugging Face 79.67B (0%). 79.7 B, counted from the tensor headers of https://huggingface.co/Qwen/Qwen3-Coder-Next at 2026-10-17T09:10:23.966Z total_parameters 79.7 B — https://huggingface.co/Qwen/Qwen3-Coder-Next counted_at 2026-10-17T09:10:23.966Z — spec_sources.verified_at for total_params_b — when this figure was counted; a refresh that keeps it does not rewrite it confidence 0.95 — spec_sources.confidence — from the tensor headers The parameters of the language-model tensors only, counted from the tensor headers: a vision or audio tower and multi-token-prediction heads are excluded, so it can be below the whole-checkpoint count the formula's cross-check quotes. It sizes the weights only when no measured file size is stored for the chosen quantization; it is not what is read per token on a mixture-of-experts model. Shown to one decimal: the underlying tensor count is exact, but more digits would claim a precision the published figure does not have (SPEC §14). 3.6 (derived) DERIVEDQwen3-Coder-Next active parameters per tokenArithmetic on observed inputs only. No tunable assumption enters it. active − input embedding table — Per-token memory read used for decode bandwidth (SPEC §8.2). 3.6 B, counted from the tensor headers of https://huggingface.co/Qwen/Qwen3-Coder-Next at 2026-10-17T09:10:23.966Z active_parameters_per_token 3.6 B — https://huggingface.co/Qwen/Qwen3-Coder-Next counted_at 2026-10-17T09:10:23.966Z — spec_sources.verified_at for active_params_b — when this figure was counted; a refresh that keeps it does not rewrite it confidence 0.95 — spec_sources.confidence — from the tensor headers The parameters read to produce one token. It sets the decode speed; on a dense model it equals the total. Shown to one decimal: the underlying tensor count is exact, but more digits would claim a precision the published figure does not have (SPEC §14). active	native 44.8 GB (observed) OBSERVEDnative file size on diskMeasured or published, passed through unchanged — not computed here. quantization file size — published at the source, passed through unchanged; nothing is computed here 44.84 GB, as recorded from https://huggingface.co/mlx-community/Qwen3-Coder-Next-nvfp4/tree/main, file 9 files quantization_file_size 44.84 GB — https://huggingface.co/mlx-community/Qwen3-Coder-Next-nvfp4/tree/main, file 9 files verified_at 2026-10-17T09:10:23.966Z — the row's own verification timestamp quantization native — mlx bits_per_weight 4.5027 bits — the quantizer's own naming, or measured from the file The on-disk size of the weights. It is not the memory the model occupies: add the KV cache and the runtime overhead (SPEC §8.1). Preferred over the bits-per-weight formula wherever it exists, because it includes embeddings and any higher-precision tensors. Observed, not derived: this app records the figure and its source, and does not re-measure it. on disk	49.2 GB (estimate) ESTIMATEQwen3-Coder-Next memory requirement at 32K contextArithmetic that also depends on a tunable assumption, listed below. required_gb = weights_gb + kv_cache_gb + overhead_gb required_gb = 44.84 + 0.81 + 3.59 = 49.24 GB weights_gb 44.844231383 GB — real file size kv_cache_gb 0.805306368 GB — kv formula overhead_gb 3.58753851064 GB — overhead fraction Real file size used instead of the params × bits formula (SPEC §8.1). 12 layers keep a full KV cache, 36 linear-attention/SSM layers keep only a constant-size state (treated as negligible). Runtime buffers, activations, and allocator slack approximated as a fraction of weights.	8 of 89 (derived) DERIVEDmachines that run Qwen3-Coder-Next at ≥ 15 tok/sArithmetic on observed inputs only. No tunable assumption enters it. machines = #{candidate rigs where the model loads AND clears the Plan's speed floor} 89 evaluated → 8 hold it in memory → 8 of those also reach ≥ 15 tok/s candidates_evaluated 89 machines — every complete rig the search could build from the priced catalog hold_it_in_memory 8 machines — required memory ≤ usable memory (SPEC §8.6) clear_the_speed_floor 8 machines — ≥ 15 tok/s The denominator is the machines this catalog can build today, not every machine that exists. A model with no shape or no sized quantization is counted on no machine at all, and says so rather than counting as zero.	SIZEABLE	read stale 3 d — past the 36 hour freshness window — Qwen3-Coder-Next: the least recent read of its data from source (shape counted Oct 17, 5:10 AM ET from the tensor headers of https://huggingface.co/Qwen/Qwen3-Coder-Next; quantization sizes read Oct 17, 5:10 AM ET); never checked upstream.
NVIDIA-Nemotron-3-Super-120B-A12BNVIDIA · MoE · reasoning · 256K ctx max	120.7 (derived) DERIVEDNVIDIA-Nemotron-3-Super-120B-A12B total parametersArithmetic on observed inputs only. No tunable assumption enters it. Σ params of language-model tensors — Cross-checked: tensor count 123.61B vs Hugging Face 123.61B (1.70e-5%). 120.7 B, counted from the tensor headers of https://huggingface.co/nvidia/NVIDIA-Nemotron-3-Super-120B-A12B-BF16 at 2026-10-17T11:30:51.565Z total_parameters 120.7 B — https://huggingface.co/nvidia/NVIDIA-Nemotron-3-Super-120B-A12B-BF16 counted_at 2026-10-17T11:30:51.565Z — spec_sources.verified_at for total_params_b — when this figure was counted; a refresh that keeps it does not rewrite it confidence 0.95 — spec_sources.confidence — from the tensor headers The parameters of the language-model tensors only, counted from the tensor headers: a vision or audio tower and multi-token-prediction heads are excluded, so it can be below the whole-checkpoint count the formula's cross-check quotes. It sizes the weights only when no measured file size is stored for the chosen quantization; it is not what is read per token on a mixture-of-experts model. Shown to one decimal: the underlying tensor count is exact, but more digits would claim a precision the published figure does not have (SPEC §14). 12.2 (derived) DERIVEDNVIDIA-Nemotron-3-Super-120B-A12B active parameters per tokenArithmetic on observed inputs only. No tunable assumption enters it. active − input embedding table — Per-token memory read used for decode bandwidth (SPEC §8.2). 12.2 B, counted from the tensor headers of https://huggingface.co/nvidia/NVIDIA-Nemotron-3-Super-120B-A12B-BF16 at 2026-10-17T11:30:51.565Z active_parameters_per_token 12.2 B — https://huggingface.co/nvidia/NVIDIA-Nemotron-3-Super-120B-A12B-BF16 counted_at 2026-10-17T11:30:51.565Z — spec_sources.verified_at for active_params_b — when this figure was counted; a refresh that keeps it does not rewrite it confidence 0.95 — spec_sources.confidence — from the tensor headers The parameters read to produce one token. It sets the decode speed; on a dense model it equals the total. Shown to one decimal: the underlying tensor count is exact, but more digits would claim a precision the published figure does not have (SPEC §14). active	4bit 68.0 GB (observed) OBSERVED4bit file size on diskMeasured or published, passed through unchanged — not computed here. quantization file size — published at the source, passed through unchanged; nothing is computed here 68.00 GB, as recorded from https://huggingface.co/mlx-community/NVIDIA-Nemotron-3-Super-120B-A12B-4bit/tree/main, file 14 files quantization_file_size 68.00 GB — https://huggingface.co/mlx-community/NVIDIA-Nemotron-3-Super-120B-A12B-4bit/tree/main, file 14 files verified_at 2026-10-17T11:30:51.565Z — the row's own verification timestamp quantization 4bit — mlx bits_per_weight 4.5083 bits — the quantizer's own naming, or measured from the file The on-disk size of the weights. It is not the memory the model occupies: add the KV cache and the runtime overhead (SPEC §8.1). Preferred over the bits-per-weight formula wherever it exists, because it includes embeddings and any higher-precision tensors. Observed, not derived: this app records the figure and its source, and does not re-measure it. on disk	73.7 GB (estimate) ESTIMATENVIDIA-Nemotron-3-Super-120B-A12B memory requirement at 32K contextArithmetic that also depends on a tunable assumption, listed below. required_gb = weights_gb + kv_cache_gb + overhead_gb required_gb = 68 + 0.27 + 5.44 = 73.71 GB weights_gb 68.000941086 GB — real file size kv_cache_gb 0.268435456 GB — kv formula overhead_gb 5.44007528688 GB — overhead fraction Real file size used instead of the params × bits formula (SPEC §8.1). 8 layers keep a full KV cache, 40 linear-attention/SSM layers keep only a constant-size state (treated as negligible). Runtime buffers, activations, and allocator slack approximated as a fraction of weights.	4 of 89 (derived) DERIVEDmachines that run NVIDIA-Nemotron-3-Super-120B-A12B at ≥ 15 tok/sArithmetic on observed inputs only. No tunable assumption enters it. machines = #{candidate rigs where the model loads AND clears the Plan's speed floor} 89 evaluated → 4 hold it in memory → 4 of those also reach ≥ 15 tok/s candidates_evaluated 89 machines — every complete rig the search could build from the priced catalog hold_it_in_memory 4 machines — required memory ≤ usable memory (SPEC §8.6) clear_the_speed_floor 4 machines — ≥ 15 tok/s The denominator is the machines this catalog can build today, not every machine that exists. A model with no shape or no sized quantization is counted on no machine at all, and says so rather than counting as zero.	SIZEABLE	read stale 3 d — past the 36 hour freshness window — NVIDIA-Nemotron-3-Super-120B-A12B: the least recent read of its data from source (shape counted Oct 17, 7:30 AM ET from the tensor headers of https://huggingface.co/nvidia/NVIDIA-Nemotron-3-Super-120B-A12B-BF16; quantization sizes read Oct 17, 7:30 AM ET); never checked upstream.
MiniMax-M2.7MiniMax · MoE · coding · 200K ctx max	228.7 (derived) DERIVEDMiniMax-M2.7 total parametersArithmetic on observed inputs only. No tunable assumption enters it. Σ params of language-model tensors — Cross-checked: tensor count 228.69B vs Hugging Face 228.69B (0%). 228.7 B, counted from the tensor headers of https://huggingface.co/MiniMaxAI/MiniMax-M2.7 at 2026-10-17T10:53:32.947Z total_parameters 228.7 B — https://huggingface.co/MiniMaxAI/MiniMax-M2.7 counted_at 2026-10-17T10:53:32.947Z — spec_sources.verified_at for total_params_b — when this figure was counted; a refresh that keeps it does not rewrite it confidence 0.95 — spec_sources.confidence — from the tensor headers The parameters of the language-model tensors only, counted from the tensor headers: a vision or audio tower and multi-token-prediction heads are excluded, so it can be below the whole-checkpoint count the formula's cross-check quotes. It sizes the weights only when no measured file size is stored for the chosen quantization; it is not what is read per token on a mixture-of-experts model. Shown to one decimal: the underlying tensor count is exact, but more digits would claim a precision the published figure does not have (SPEC §14). 10.4 (derived) DERIVEDMiniMax-M2.7 active parameters per tokenArithmetic on observed inputs only. No tunable assumption enters it. active − input embedding table — Per-token memory read used for decode bandwidth (SPEC §8.2). 10.4 B, counted from the tensor headers of https://huggingface.co/MiniMaxAI/MiniMax-M2.7 at 2026-10-17T10:53:32.947Z active_parameters_per_token 10.4 B — https://huggingface.co/MiniMaxAI/MiniMax-M2.7 counted_at 2026-10-17T10:53:32.947Z — spec_sources.verified_at for active_params_b — when this figure was counted; a refresh that keeps it does not rewrite it confidence 0.95 — spec_sources.confidence — from the tensor headers The parameters read to produce one token. It sets the decode speed; on a dense model it equals the total. Shown to one decimal: the underlying tensor count is exact, but more digits would claim a precision the published figure does not have (SPEC §14). active	4bit 129 GB (observed) OBSERVED4bit file size on diskMeasured or published, passed through unchanged — not computed here. quantization file size — published at the source, passed through unchanged; nothing is computed here 128.67 GB, as recorded from https://huggingface.co/mlx-community/MiniMax-M2.7-4bit/tree/main, file 27 files quantization_file_size 128.67 GB — https://huggingface.co/mlx-community/MiniMax-M2.7-4bit/tree/main, file 27 files verified_at 2026-10-17T10:53:32.947Z — the row's own verification timestamp quantization 4bit — mlx bits_per_weight 4.501 bits — the quantizer's own naming, or measured from the file The on-disk size of the weights. It is not the memory the model occupies: add the KV cache and the runtime overhead (SPEC §8.1). Preferred over the bits-per-weight formula wherever it exists, because it includes embeddings and any higher-precision tensors. Observed, not derived: this app records the figure and its source, and does not re-measure it. on disk	147 GB (estimate) ESTIMATEMiniMax-M2.7 memory requirement at 32K contextArithmetic that also depends on a tunable assumption, listed below. required_gb = weights_gb + kv_cache_gb + overhead_gb required_gb = 128.67 + 8.32 + 10.29 = 147.28 GB weights_gb 128.66688798 GB — real file size kv_cache_gb 8.321499136 GB — kv formula overhead_gb 10.2933510384 GB — overhead fraction Real file size used instead of the params × bits formula (SPEC §8.1). 62 layers keep a full KV cache. Runtime buffers, activations, and allocator slack approximated as a fraction of weights.	3 of 89 (derived) DERIVEDmachines that run MiniMax-M2.7 at ≥ 15 tok/sArithmetic on observed inputs only. No tunable assumption enters it. machines = #{candidate rigs where the model loads AND clears the Plan's speed floor} 89 evaluated → 3 hold it in memory → 3 of those also reach ≥ 15 tok/s candidates_evaluated 89 machines — every complete rig the search could build from the priced catalog hold_it_in_memory 3 machines — required memory ≤ usable memory (SPEC §8.6) clear_the_speed_floor 3 machines — ≥ 15 tok/s The denominator is the machines this catalog can build today, not every machine that exists. A model with no shape or no sized quantization is counted on no machine at all, and says so rather than counting as zero.	SIZEABLE	read stale 3 d — past the 36 hour freshness window — MiniMax-M2.7: the least recent read of its data from source (shape counted Oct 17, 6:53 AM ET from the tensor headers of https://huggingface.co/MiniMaxAI/MiniMax-M2.7; quantization sizes read Oct 17, 6:53 AM ET); never checked upstream.
Kimi-K2.7-CodeMoonshot AI · MoE · coding · 256K ctx max	1,026.4 (derived) DERIVEDKimi-K2.7-Code total parametersArithmetic on observed inputs only. No tunable assumption enters it. Σ params of language-model tensors — Cross-checked: tensor count 1,026.88B vs Hugging Face 1,026.88B (0%). 1026.4 B, counted from the tensor headers of https://huggingface.co/moonshotai/Kimi-K2.7-Code at 2026-10-17T10:18:54.827Z total_parameters 1026.4 B — https://huggingface.co/moonshotai/Kimi-K2.7-Code counted_at 2026-10-17T10:18:54.827Z — spec_sources.verified_at for total_params_b — when this figure was counted; a refresh that keeps it does not rewrite it confidence 0.95 — spec_sources.confidence — from the tensor headers The parameters of the language-model tensors only, counted from the tensor headers: a vision or audio tower and multi-token-prediction heads are excluded, so it can be below the whole-checkpoint count the formula's cross-check quotes. It sizes the weights only when no measured file size is stored for the chosen quantization; it is not what is read per token on a mixture-of-experts model. Shown to one decimal: the underlying tensor count is exact, but more digits would claim a precision the published figure does not have (SPEC §14). 31.7 (derived) DERIVEDKimi-K2.7-Code active parameters per tokenArithmetic on observed inputs only. No tunable assumption enters it. active − input embedding table — Per-token memory read used for decode bandwidth (SPEC §8.2). 31.7 B, counted from the tensor headers of https://huggingface.co/moonshotai/Kimi-K2.7-Code at 2026-10-17T10:18:54.827Z active_parameters_per_token 31.7 B — https://huggingface.co/moonshotai/Kimi-K2.7-Code counted_at 2026-10-17T10:18:54.827Z — spec_sources.verified_at for active_params_b — when this figure was counted; a refresh that keeps it does not rewrite it confidence 0.95 — spec_sources.confidence — from the tensor headers The parameters read to produce one token. It sets the decode speed; on a dense model it equals the total. Shown to one decimal: the underlying tensor count is exact, but more digits would claim a precision the published figure does not have (SPEC §14). active	UD-Q4_K_XL 584 GB (observed) OBSERVEDUD-Q4_K_XL file size on diskMeasured or published, passed through unchanged — not computed here. quantization file size — published at the source, passed through unchanged; nothing is computed here 583.71 GB, as recorded from https://huggingface.co/unsloth/Kimi-K2.7-Code-GGUF/tree/main, file 14 files quantization_file_size 583.71 GB — https://huggingface.co/unsloth/Kimi-K2.7-Code-GGUF/tree/main, file 14 files verified_at 2026-10-17T10:18:54.827Z — the row's own verification timestamp quantization UD-Q4_K_XL — gguf bits_per_weight 4.5495 bits — the quantizer's own naming, or measured from the file The on-disk size of the weights. It is not the memory the model occupies: add the KV cache and the runtime overhead (SPEC §8.1). Preferred over the bits-per-weight formula wherever it exists, because it includes embeddings and any higher-precision tensors. Observed, not derived: this app records the figure and its source, and does not re-measure it. on disk	633 GB (estimate) ESTIMATEKimi-K2.7-Code memory requirement at 32K contextArithmetic that also depends on a tunable assumption, listed below. required_gb = weights_gb + kv_cache_gb + overhead_gb required_gb = 583.71 + 2.3 + 46.7 = 632.71 GB weights_gb 583.71087552 GB — real file size kv_cache_gb 2.302672896 GB — kv formula overhead_gb 46.69687004159999 GB — overhead fraction Real file size used instead of the params × bits formula (SPEC §8.1). 61 layers keep a full KV cache. Runtime buffers, activations, and allocator slack approximated as a fraction of weights.	1 of 89 (derived) DERIVEDmachines that run Kimi-K2.7-Code at ≥ 15 tok/sArithmetic on observed inputs only. No tunable assumption enters it. machines = #{candidate rigs where the model loads AND clears the Plan's speed floor} 89 evaluated → 19 hold it in memory → 1 of those also reach ≥ 15 tok/s candidates_evaluated 89 machines — every complete rig the search could build from the priced catalog hold_it_in_memory 19 machines — required memory ≤ usable memory (SPEC §8.6) clear_the_speed_floor 1 machines — ≥ 15 tok/s The denominator is the machines this catalog can build today, not every machine that exists. A model with no shape or no sized quantization is counted on no machine at all, and says so rather than counting as zero.	SIZEABLE	read stale 3 d — past the 36 hour freshness window — Kimi-K2.7-Code: the least recent read of its data from source (shape counted Oct 17, 6:18 AM ET from the tensor headers of https://huggingface.co/moonshotai/Kimi-K2.7-Code; quantization sizes read Oct 17, 6:18 AM ET); never checked upstream.
MiniMax-M3MiniMax · MoE · coding, vision · 1M ctx max	426.4 (derived) DERIVEDMiniMax-M3 total parametersArithmetic on observed inputs only. No tunable assumption enters it. Σ params of language-model tensors — Cross-checked: tensor count 427.04B vs Hugging Face 427.04B (0%). 426.4 B, counted from the tensor headers of https://huggingface.co/MiniMaxAI/MiniMax-M3 at 2026-10-17T10:40:15.227Z total_parameters 426.4 B — https://huggingface.co/MiniMaxAI/MiniMax-M3 counted_at 2026-10-17T10:40:15.227Z — spec_sources.verified_at for total_params_b — when this figure was counted; a refresh that keeps it does not rewrite it confidence 0.95 — spec_sources.confidence — from the tensor headers The parameters of the language-model tensors only, counted from the tensor headers: a vision or audio tower and multi-token-prediction heads are excluded, so it can be below the whole-checkpoint count the formula's cross-check quotes. It sizes the weights only when no measured file size is stored for the chosen quantization; it is not what is read per token on a mixture-of-experts model. Shown to one decimal: the underlying tensor count is exact, but more digits would claim a precision the published figure does not have (SPEC §14). 24.9 (derived) DERIVEDMiniMax-M3 active parameters per tokenArithmetic on observed inputs only. No tunable assumption enters it. active − input embedding table — Per-token memory read used for decode bandwidth (SPEC §8.2). 24.9 B, counted from the tensor headers of https://huggingface.co/MiniMaxAI/MiniMax-M3 at 2026-10-17T10:40:15.227Z active_parameters_per_token 24.9 B — https://huggingface.co/MiniMaxAI/MiniMax-M3 counted_at 2026-10-17T10:40:15.227Z — spec_sources.verified_at for active_params_b — when this figure was counted; a refresh that keeps it does not rewrite it confidence 0.95 — spec_sources.confidence — from the tensor headers The parameters read to produce one token. It sets the decode speed; on a dense model it equals the total. Shown to one decimal: the underlying tensor count is exact, but more digits would claim a precision the published figure does not have (SPEC §14). active	4bit 241 GB (observed) OBSERVED4bit file size on diskMeasured or published, passed through unchanged — not computed here. quantization file size — published at the source, passed through unchanged; nothing is computed here 241.48 GB, as recorded from https://huggingface.co/mlx-community/MiniMax-M3-4bit/tree/main, file 58 files quantization_file_size 241.48 GB — https://huggingface.co/mlx-community/MiniMax-M3-4bit/tree/main, file 58 files verified_at 2026-10-17T10:40:15.227Z — the row's own verification timestamp quantization 4bit — mlx bits_per_weight 4.531 bits — the quantizer's own naming, or measured from the file The on-disk size of the weights. It is not the memory the model occupies: add the KV cache and the runtime overhead (SPEC §8.1). Preferred over the bits-per-weight formula wherever it exists, because it includes embeddings and any higher-precision tensors. Observed, not derived: this app records the figure and its source, and does not re-measure it. on disk	265 GB (estimate) ESTIMATEMiniMax-M3 memory requirement at 32K contextArithmetic that also depends on a tunable assumption, listed below. required_gb = weights_gb + kv_cache_gb + overhead_gb required_gb = 241.48 + 4.03 + 19.32 = 264.83 GB weights_gb 241.481031547 GB — real file size kv_cache_gb 4.02653184 GB — kv formula overhead_gb 19.31848252376 GB — overhead fraction Real file size used instead of the params × bits formula (SPEC §8.1). 60 layers keep a full KV cache. Runtime buffers, activations, and allocator slack approximated as a fraction of weights.	1 of 89 (derived) DERIVEDmachines that run MiniMax-M3 at ≥ 15 tok/sArithmetic on observed inputs only. No tunable assumption enters it. machines = #{candidate rigs where the model loads AND clears the Plan's speed floor} 89 evaluated → 1 hold it in memory → 1 of those also reach ≥ 15 tok/s candidates_evaluated 89 machines — every complete rig the search could build from the priced catalog hold_it_in_memory 1 machines — required memory ≤ usable memory (SPEC §8.6) clear_the_speed_floor 1 machines — ≥ 15 tok/s The denominator is the machines this catalog can build today, not every machine that exists. A model with no shape or no sized quantization is counted on no machine at all, and says so rather than counting as zero.	SIZEABLE	read stale 3 d — past the 36 hour freshness window — MiniMax-M3: the least recent read of its data from source (shape counted Oct 17, 6:40 AM ET from the tensor headers of https://huggingface.co/MiniMaxAI/MiniMax-M3; quantization sizes read Oct 17, 6:40 AM ET); never checked upstream.
Kimi K3Moonshot AI · MoE · reasoning, coding, vision · 1M ctx max	2,779.5 (derived) DERIVEDKimi K3 total parametersArithmetic on observed inputs only. No tunable assumption enters it. Σ params of language-model tensors — Cross-checked: tensor count 2,779.93B vs Hugging Face 2,779.93B (0%). 2779.5 B, counted from the tensor headers of https://huggingface.co/moonshotai/Kimi-K3 at 2026-10-17T09:57:46.951Z total_parameters 2779.5 B — https://huggingface.co/moonshotai/Kimi-K3 counted_at 2026-10-17T09:57:46.951Z — spec_sources.verified_at for total_params_b — when this figure was counted; a refresh that keeps it does not rewrite it confidence 0.95 — spec_sources.confidence — from the tensor headers The parameters of the language-model tensors only, counted from the tensor headers: a vision or audio tower and multi-token-prediction heads are excluded, so it can be below the whole-checkpoint count the formula's cross-check quotes. It sizes the weights only when no measured file size is stored for the chosen quantization; it is not what is read per token on a mixture-of-experts model. Shown to one decimal: the underlying tensor count is exact, but more digits would claim a precision the published figure does not have (SPEC §14). 207.5 (derived) DERIVEDKimi K3 active parameters per tokenArithmetic on observed inputs only. No tunable assumption enters it. active − input embedding table — Per-token memory read used for decode bandwidth (SPEC §8.2). 207.5 B, counted from the tensor headers of https://huggingface.co/moonshotai/Kimi-K3 at 2026-10-17T09:57:46.951Z active_parameters_per_token 207.5 B — https://huggingface.co/moonshotai/Kimi-K3 counted_at 2026-10-17T09:57:46.951Z — spec_sources.verified_at for active_params_b — when this figure was counted; a refresh that keeps it does not rewrite it confidence 0.95 — spec_sources.confidence — from the tensor headers The parameters read to produce one token. It sets the decode speed; on a dense model it equals the total. Shown to one decimal: the underlying tensor count is exact, but more digits would claim a precision the published figure does not have (SPEC §14). active	UD-Q8_K_XL 1,561 GB (observed) OBSERVEDUD-Q8_K_XL file size on diskMeasured or published, passed through unchanged — not computed here. quantization file size — published at the source, passed through unchanged; nothing is computed here 1561.16 GB, as recorded from https://huggingface.co/unsloth/Kimi-K3-GGUF/tree/main, file 34 files quantization_file_size 1561.16 GB — https://huggingface.co/unsloth/Kimi-K3-GGUF/tree/main, file 34 files verified_at 2026-10-17T09:57:46.951Z — the row's own verification timestamp quantization UD-Q8_K_XL — gguf bits_per_weight 4.4934 bits — the quantizer's own naming, or measured from the file The on-disk size of the weights. It is not the memory the model occupies: add the KV cache and the runtime overhead (SPEC §8.1). Preferred over the bits-per-weight formula wherever it exists, because it includes embeddings and any higher-precision tensors. Observed, not derived: this app records the figure and its source, and does not re-measure it. on disk	1,687 GB (estimate) ESTIMATEKimi K3 memory requirement at 32K contextArithmetic that also depends on a tunable assumption, listed below. required_gb = weights_gb + kv_cache_gb + overhead_gb required_gb = 1,561.16 + 0.91 + 124.89 = 1,686.96 GB weights_gb 1561.157884384 GB — real file size kv_cache_gb 0.905969664 GB — kv formula overhead_gb 124.89263075072 GB — overhead fraction Real file size used instead of the params × bits formula (SPEC §8.1). 24 layers keep a full KV cache, 69 linear-attention/SSM layers keep only a constant-size state (treated as negligible). Runtime buffers, activations, and allocator slack approximated as a fraction of weights.	none: 0 of 89 (derived) DERIVEDmachines that run Kimi K3 at ≥ 15 tok/sArithmetic on observed inputs only. No tunable assumption enters it. machines = #{candidate rigs where the model loads AND clears the Plan's speed floor} 89 evaluated → 0 hold it in memory → 0 of those also reach ≥ 15 tok/s candidates_evaluated 89 machines — every complete rig the search could build from the priced catalog hold_it_in_memory 0 machines — required memory ≤ usable memory (SPEC §8.6) clear_the_speed_floor 0 machines — ≥ 15 tok/s The denominator is the machines this catalog can build today, not every machine that exists. A model with no shape or no sized quantization is counted on no machine at all, and says so rather than counting as zero.	SIZEABLE	read stale 3 d — past the 36 hour freshness window — Kimi K3: the least recent read of its data from source (shape counted Oct 17, 5:57 AM ET from the tensor headers of https://huggingface.co/moonshotai/Kimi-K3; quantization sizes read Oct 17, 5:57 AM ET); never checked upstream.
Qwen3.8-2.4T-A95B (open Qwen3.8-Max)Alibaba Qwen · MoE · reasoning, coding · 256K ctx max	2,419.8 (derived) DERIVEDQwen3.8-2.4T-A95B (open Qwen3.8-Max) total parametersArithmetic on observed inputs only. No tunable assumption enters it. Σ params of language-model tensors — Cross-checked: tensor count 2,446.18B vs Hugging Face 2,446.18B (0%). 2419.8 B, counted from the tensor headers of https://huggingface.co/Qwen/Qwen3.8-2.4T-A95B at 2026-10-18T07:40:56.336Z total_parameters 2419.8 B — https://huggingface.co/Qwen/Qwen3.8-2.4T-A95B counted_at 2026-10-18T07:40:56.336Z — spec_sources.verified_at for total_params_b — when this figure was counted; a refresh that keeps it does not rewrite it confidence 0.95 — spec_sources.confidence — from the tensor headers The parameters of the language-model tensors only, counted from the tensor headers: a vision or audio tower and multi-token-prediction heads are excluded, so it can be below the whole-checkpoint count the formula's cross-check quotes. It sizes the weights only when no measured file size is stored for the chosen quantization; it is not what is read per token on a mixture-of-experts model. Shown to one decimal: the underlying tensor count is exact, but more digits would claim a precision the published figure does not have (SPEC §14). 93.3 (derived) DERIVEDQwen3.8-2.4T-A95B (open Qwen3.8-Max) active parameters per tokenArithmetic on observed inputs only. No tunable assumption enters it. active − input embedding table — Per-token memory read used for decode bandwidth (SPEC §8.2). 93.3 B, counted from the tensor headers of https://huggingface.co/Qwen/Qwen3.8-2.4T-A95B at 2026-10-18T07:40:56.336Z active_parameters_per_token 93.3 B — https://huggingface.co/Qwen/Qwen3.8-2.4T-A95B counted_at 2026-10-18T07:40:56.336Z — spec_sources.verified_at for active_params_b — when this figure was counted; a refresh that keeps it does not rewrite it confidence 0.95 — spec_sources.confidence — from the tensor headers The parameters read to produce one token. It sets the decode speed; on a dense model it equals the total. Shown to one decimal: the underlying tensor count is exact, but more digits would claim a precision the published figure does not have (SPEC §14). active	UD-IQ4_XS 1,311 GB (observed) OBSERVEDUD-IQ4_XS file size on diskMeasured or published, passed through unchanged — not computed here. quantization file size — published at the source, passed through unchanged; nothing is computed here 1310.88 GB, as recorded from https://huggingface.co/unsloth/Qwen3.8-2.4T-A95B-GGUF/tree/main, file 29 files quantization_file_size 1310.88 GB — https://huggingface.co/unsloth/Qwen3.8-2.4T-A95B-GGUF/tree/main, file 29 files verified_at 2026-10-18T07:40:56.336Z — the row's own verification timestamp quantization UD-IQ4_XS — gguf bits_per_weight 4.3338 bits — the quantizer's own naming, or measured from the file The on-disk size of the weights. It is not the memory the model occupies: add the KV cache and the runtime overhead (SPEC §8.1). Preferred over the bits-per-weight formula wherever it exists, because it includes embeddings and any higher-precision tensors. Observed, not derived: this app records the figure and its source, and does not re-measure it. on disk	1,419 GB (estimate) ESTIMATEQwen3.8-2.4T-A95B (open Qwen3.8-Max) memory requirement at 32K contextArithmetic that also depends on a tunable assumption, listed below. required_gb = weights_gb + kv_cache_gb + overhead_gb required_gb = 1,310.88 + 3.09 + 104.87 = 1,418.83 GB weights_gb 1310.876469952 GB — real file size kv_cache_gb 3.087007744 GB — kv formula overhead_gb 104.87011759616 GB — overhead fraction Real file size used instead of the params × bits formula (SPEC §8.1). 23 layers keep a full KV cache, 69 linear-attention/SSM layers keep only a constant-size state (treated as negligible). Runtime buffers, activations, and allocator slack approximated as a fraction of weights.	none: 0 of 89 (derived) DERIVEDmachines that run Qwen3.8-2.4T-A95B (open Qwen3.8-Max) at ≥ 15 tok/sArithmetic on observed inputs only. No tunable assumption enters it. machines = #{candidate rigs where the model loads AND clears the Plan's speed floor} 89 evaluated → 0 hold it in memory → 0 of those also reach ≥ 15 tok/s candidates_evaluated 89 machines — every complete rig the search could build from the priced catalog hold_it_in_memory 0 machines — required memory ≤ usable memory (SPEC §8.6) clear_the_speed_floor 0 machines — ≥ 15 tok/s The denominator is the machines this catalog can build today, not every machine that exists. A model with no shape or no sized quantization is counted on no machine at all, and says so rather than counting as zero.	SIZEABLE	read stale 2 d — past the 36 hour freshness window — Qwen3.8-2.4T-A95B (open Qwen3.8-Max): the least recent read of its data from source (shape counted Oct 18, 3:40 AM ET from the tensor headers of https://huggingface.co/Qwen/Qwen3.8-2.4T-A95B; quantization sizes read Oct 18, 3:40 AM ET); no upstream check after it.
In the saved Plan: DeepSeek-V4-Pro-0813DeepSeek · MoE · reasoning, coding · 1M ctx max	1,573.0 (derived) DERIVEDDeepSeek-V4-Pro-0813 total parametersArithmetic on observed inputs only. No tunable assumption enters it. Σ params of language-model tensors — Cross-checked: tensor count 1,650.5B vs Hugging Face 1,650.5B (-1.41e-4%). 1573.0 B, counted from the tensor headers of https://huggingface.co/deepseek-ai/DeepSeek-V4-Pro-0813 at 2026-10-17T09:20:49.747Z total_parameters 1573.0 B — https://huggingface.co/deepseek-ai/DeepSeek-V4-Pro-0813 counted_at 2026-10-17T09:20:49.747Z — spec_sources.verified_at for total_params_b — when this figure was counted; a refresh that keeps it does not rewrite it confidence 0.95 — spec_sources.confidence — from the tensor headers The parameters of the language-model tensors only, counted from the tensor headers: a vision or audio tower and multi-token-prediction heads are excluded, so it can be below the whole-checkpoint count the formula's cross-check quotes. It sizes the weights only when no measured file size is stored for the chosen quantization; it is not what is read per token on a mixture-of-experts model. Shown to one decimal: the underlying tensor count is exact, but more digits would claim a precision the published figure does not have (SPEC §14). 48.9 (derived) DERIVEDDeepSeek-V4-Pro-0813 active parameters per tokenArithmetic on observed inputs only. No tunable assumption enters it. active − input embedding table — Per-token memory read used for decode bandwidth (SPEC §8.2). 48.9 B, counted from the tensor headers of https://huggingface.co/deepseek-ai/DeepSeek-V4-Pro-0813 at 2026-10-17T09:20:49.747Z active_parameters_per_token 48.9 B — https://huggingface.co/deepseek-ai/DeepSeek-V4-Pro-0813 counted_at 2026-10-17T09:20:49.747Z — spec_sources.verified_at for active_params_b — when this figure was counted; a refresh that keeps it does not rewrite it confidence 0.95 — spec_sources.confidence — from the tensor headers The parameters read to produce one token. It sets the decode speed; on a dense model it equals the total. Shown to one decimal: the underlying tensor count is exact, but more digits would claim a precision the published figure does not have (SPEC §14). active	native 893 GB (observed) OBSERVEDnative file size on diskMeasured or published, passed through unchanged — not computed here. quantization file size — published at the source, passed through unchanged; nothing is computed here 892.74 GB, as recorded from https://huggingface.co/unsloth/DeepSeek-V4-Pro-0813/tree/main, file 66 files quantization_file_size 892.74 GB — https://huggingface.co/unsloth/DeepSeek-V4-Pro-0813/tree/main, file 66 files verified_at 2026-10-17T09:20:49.747Z — the row's own verification timestamp quantization native — safetensors bits_per_weight 4.5403 bits — the quantizer's own naming, or measured from the file The on-disk size of the weights. It is not the memory the model occupies: add the KV cache and the runtime overhead (SPEC §8.1). Preferred over the bits-per-weight formula wherever it exists, because it includes embeddings and any higher-precision tensors. Observed, not derived: this app records the figure and its source, and does not re-measure it. on disk	966 GB (estimate) ESTIMATEDeepSeek-V4-Pro-0813 memory requirement at 32K contextArithmetic that also depends on a tunable assumption, listed below. required_gb = weights_gb + kv_cache_gb + overhead_gb required_gb = 892.74 + 2.05 + 71.42 = 966.21 GB weights_gb 892.74432288 GB — real file size kv_cache_gb 2.046820352 GB — kv formula overhead_gb 71.4195458304 GB — overhead fraction Real file size used instead of the params × bits formula (SPEC §8.1). 61 layers keep a full KV cache. KV width inferred from a single latent 'wkv' projection (unrecognized attention family); any KV compression the runtime applies is not modeled, so the cache is likely overstated. Runtime buffers, activations, and allocator slack approximated as a fraction of weights.	none: 0 of 89 (derived) DERIVEDmachines that run DeepSeek-V4-Pro-0813 at ≥ 15 tok/sArithmetic on observed inputs only. No tunable assumption enters it. machines = #{candidate rigs where the model loads AND clears the Plan's speed floor} 89 evaluated → 9 hold it in memory → 0 of those also reach ≥ 15 tok/s candidates_evaluated 89 machines — every complete rig the search could build from the priced catalog hold_it_in_memory 9 machines — required memory ≤ usable memory (SPEC §8.6) clear_the_speed_floor 0 machines — ≥ 15 tok/s The denominator is the machines this catalog can build today, not every machine that exists. A model with no shape or no sized quantization is counted on no machine at all, and says so rather than counting as zero.	SIZEABLE	read stale 3 d — past the 36 hour freshness window — DeepSeek-V4-Pro-0813: the least recent read of its data from source (shape counted Oct 17, 5:20 AM ET from the tensor headers of https://huggingface.co/deepseek-ai/DeepSeek-V4-Pro-0813; quantization sizes read Oct 17, 5:20 AM ET); never checked upstream.
In the saved Plan: GLM-5.3Z.ai (zai-org) · MoE · coding, reasoning · 1M ctx max	743.4 (derived) DERIVEDGLM-5.3 total parametersArithmetic on observed inputs only. No tunable assumption enters it. Σ params of language-model tensors — Cross-checked: tensor count 753.33B vs Hugging Face 753.33B (0%). 743.4 B, counted from the tensor headers of https://huggingface.co/zai-org/GLM-5.3 at 2026-10-18T07:21:53.314Z total_parameters 743.4 B — https://huggingface.co/zai-org/GLM-5.3 counted_at 2026-10-18T07:21:53.314Z — spec_sources.verified_at for total_params_b — when this figure was counted; a refresh that keeps it does not rewrite it confidence 0.95 — spec_sources.confidence — from the tensor headers The parameters of the language-model tensors only, counted from the tensor headers: a vision or audio tower and multi-token-prediction heads are excluded, so it can be below the whole-checkpoint count the formula's cross-check quotes. It sizes the weights only when no measured file size is stored for the chosen quantization; it is not what is read per token on a mixture-of-experts model. Shown to one decimal: the underlying tensor count is exact, but more digits would claim a precision the published figure does not have (SPEC §14). 40.3 (derived) DERIVEDGLM-5.3 active parameters per tokenArithmetic on observed inputs only. No tunable assumption enters it. active − input embedding table — Per-token memory read used for decode bandwidth (SPEC §8.2). 40.3 B, counted from the tensor headers of https://huggingface.co/zai-org/GLM-5.3 at 2026-10-18T07:21:53.314Z active_parameters_per_token 40.3 B — https://huggingface.co/zai-org/GLM-5.3 counted_at 2026-10-18T07:21:53.314Z — spec_sources.verified_at for active_params_b — when this figure was counted; a refresh that keeps it does not rewrite it confidence 0.95 — spec_sources.confidence — from the tensor headers The parameters read to produce one token. It sets the decode speed; on a dense model it equals the total. Shown to one decimal: the underlying tensor count is exact, but more digits would claim a precision the published figure does not have (SPEC §14). active	UD-Q4_K_XL 467 GB (observed) OBSERVEDUD-Q4_K_XL file size on diskMeasured or published, passed through unchanged — not computed here. quantization file size — published at the source, passed through unchanged; nothing is computed here 467.29 GB, as recorded from https://huggingface.co/unsloth/GLM-5.3-GGUF/tree/main, file 11 files quantization_file_size 467.29 GB — https://huggingface.co/unsloth/GLM-5.3-GGUF/tree/main, file 11 files verified_at 2026-10-18T07:21:53.314Z — the row's own verification timestamp quantization UD-Q4_K_XL — gguf bits_per_weight 5.0288 bits — the quantizer's own naming, or measured from the file The on-disk size of the weights. It is not the memory the model occupies: add the KV cache and the runtime overhead (SPEC §8.1). Preferred over the bits-per-weight formula wherever it exists, because it includes embeddings and any higher-precision tensors. Observed, not derived: this app records the figure and its source, and does not re-measure it. on disk	508 GB (estimate) ESTIMATEGLM-5.3 memory requirement at 32K contextArithmetic that also depends on a tunable assumption, listed below. required_gb = weights_gb + kv_cache_gb + overhead_gb required_gb = 467.29 + 2.94 + 37.38 = 507.62 GB weights_gb 467.289116837 GB — real file size kv_cache_gb 2.944401408 GB — kv formula overhead_gb 37.38312934696 GB — overhead fraction Real file size used instead of the params × bits formula (SPEC §8.1). 78 layers keep a full KV cache. Runtime buffers, activations, and allocator slack approximated as a fraction of weights.	none: 0 of 89 (derived) DERIVEDmachines that run GLM-5.3 at ≥ 15 tok/sArithmetic on observed inputs only. No tunable assumption enters it. machines = #{candidate rigs where the model loads AND clears the Plan's speed floor} 89 evaluated → 26 hold it in memory → 0 of those also reach ≥ 15 tok/s candidates_evaluated 89 machines — every complete rig the search could build from the priced catalog hold_it_in_memory 26 machines — required memory ≤ usable memory (SPEC §8.6) clear_the_speed_floor 0 machines — ≥ 15 tok/s The denominator is the machines this catalog can build today, not every machine that exists. A model with no shape or no sized quantization is counted on no machine at all, and says so rather than counting as zero.	SIZEABLE	read stale 3 d — past the 36 hour freshness window — GLM-5.3: the least recent read of its data from source (shape counted Oct 18, 3:21 AM ET from the tensor headers of https://huggingface.co/zai-org/GLM-5.3; quantization sizes read Oct 18, 3:21 AM ET); no upstream check after it.
Step-3.7-FlashStepFun · MoE · vision · 256K ctx max	197.0 (derived) DERIVEDStep-3.7-Flash total parametersArithmetic on observed inputs only. No tunable assumption enters it. Σ params of language-model tensors — Cross-checked: tensor count 201.37B vs Hugging Face 201.37B (0%). 197.0 B, counted from the tensor headers of https://huggingface.co/stepfun-ai/Step-3.7-Flash at 2026-10-17T11:43:09.838Z total_parameters 197.0 B — https://huggingface.co/stepfun-ai/Step-3.7-Flash counted_at 2026-10-17T11:43:09.838Z — spec_sources.verified_at for total_params_b — when this figure was counted; a refresh that keeps it does not rewrite it confidence 0.95 — spec_sources.confidence — from the tensor headers The parameters of the language-model tensors only, counted from the tensor headers: a vision or audio tower and multi-token-prediction heads are excluded, so it can be below the whole-checkpoint count the formula's cross-check quotes. It sizes the weights only when no measured file size is stored for the chosen quantization; it is not what is read per token on a mixture-of-experts model. Shown to one decimal: the underlying tensor count is exact, but more digits would claim a precision the published figure does not have (SPEC §14). 196.5 (derived) DERIVEDStep-3.7-Flash active parameters per tokenArithmetic on observed inputs only. No tunable assumption enters it. active − input embedding table — Per-token memory read used for decode bandwidth (SPEC §8.2). 196.5 B, counted from the tensor headers of https://huggingface.co/stepfun-ai/Step-3.7-Flash at 2026-10-17T11:43:09.838Z active_parameters_per_token 196.5 B — https://huggingface.co/stepfun-ai/Step-3.7-Flash counted_at 2026-10-17T11:43:09.838Z — spec_sources.verified_at for active_params_b — when this figure was counted; a refresh that keeps it does not rewrite it confidence 0.95 — spec_sources.confidence — from the tensor headers The parameters read to produce one token. It sets the decode speed; on a dense model it equals the total. Shown to one decimal: the underlying tensor count is exact, but more digits would claim a precision the published figure does not have (SPEC §14). active	4bit 111 GB (observed) OBSERVED4bit file size on diskMeasured or published, passed through unchanged — not computed here. quantization file size — published at the source, passed through unchanged; nothing is computed here 110.81 GB, as recorded from https://huggingface.co/mlx-community/Step-3.7-Flash-4bit/tree/main, file 22 files quantization_file_size 110.81 GB — https://huggingface.co/mlx-community/Step-3.7-Flash-4bit/tree/main, file 22 files verified_at 2026-10-17T11:43:09.838Z — the row's own verification timestamp quantization 4bit — mlx bits_per_weight 4.5005 bits — the quantizer's own naming, or measured from the file The on-disk size of the weights. It is not the memory the model occupies: add the KV cache and the runtime overhead (SPEC §8.1). Preferred over the bits-per-weight formula wherever it exists, because it includes embeddings and any higher-precision tensors. Observed, not derived: this app records the figure and its source, and does not re-measure it. on disk	121 GB (estimate) ESTIMATEStep-3.7-Flash memory requirement at 32K contextArithmetic that also depends on a tunable assumption, listed below. required_gb = weights_gb + kv_cache_gb + overhead_gb required_gb = 110.81 + 1.68 + 8.87 = 121.36 GB weights_gb 110.813433037 GB — real file size kv_cache_gb 1.679818752 GB — kv formula overhead_gb 8.86507464296 GB — overhead fraction Real file size used instead of the params × bits formula (SPEC §8.1). 12 layers keep a full KV cache, 33 sliding-window layers are capped at 512 tokens. Runtime buffers, activations, and allocator slack approximated as a fraction of weights.	none: 0 of 89 (derived) DERIVEDmachines that run Step-3.7-Flash at ≥ 15 tok/sArithmetic on observed inputs only. No tunable assumption enters it. machines = #{candidate rigs where the model loads AND clears the Plan's speed floor} 89 evaluated → 3 hold it in memory → 0 of those also reach ≥ 15 tok/s candidates_evaluated 89 machines — every complete rig the search could build from the priced catalog hold_it_in_memory 3 machines — required memory ≤ usable memory (SPEC §8.6) clear_the_speed_floor 0 machines — ≥ 15 tok/s The denominator is the machines this catalog can build today, not every machine that exists. A model with no shape or no sized quantization is counted on no machine at all, and says so rather than counting as zero.	SIZEABLE	read stale 3 d — past the 36 hour freshness window — Step-3.7-Flash: the least recent read of its data from source (shape counted Oct 17, 7:43 AM ET from the tensor headers of https://huggingface.co/stepfun-ai/Step-3.7-Flash; quantization sizes read Oct 17, 7:43 AM ET); never checked upstream.
Canary-Qwen 2.5BNVIDIA · dense · speech to text	2.6 (derived) DERIVEDCanary-Qwen 2.5B total parametersArithmetic on observed inputs only. No tunable assumption enters it. Σ params of language-model tensors — Cross-checked: tensor count 2.56B vs Hugging Face 2.56B (1.25e-6%). 2.6 B, counted from the tensor headers of https://huggingface.co/nvidia/canary-qwen-2.5b at 2026-10-17T11:52:20.219Z total_parameters 2.6 B — https://huggingface.co/nvidia/canary-qwen-2.5b counted_at 2026-10-17T11:52:20.219Z — spec_sources.verified_at for total_params_b — when this figure was counted; a refresh that keeps it does not rewrite it confidence 0.95 — spec_sources.confidence — from the tensor headers The parameters of the language-model tensors only, counted from the tensor headers: a vision or audio tower and multi-token-prediction heads are excluded, so it can be below the whole-checkpoint count the formula's cross-check quotes. It sizes the weights only when no measured file size is stored for the chosen quantization; it is not what is read per token on a mixture-of-experts model. Shown to one decimal: the underlying tensor count is exact, but more digits would claim a precision the published figure does not have (SPEC §14). 2.2 (derived) DERIVEDCanary-Qwen 2.5B active parameters per tokenArithmetic on observed inputs only. No tunable assumption enters it. active − input embedding table — Per-token memory read used for decode bandwidth (SPEC §8.2). 2.2 B, counted from the tensor headers of https://huggingface.co/nvidia/canary-qwen-2.5b at 2026-10-17T11:52:20.219Z active_parameters_per_token 2.2 B — https://huggingface.co/nvidia/canary-qwen-2.5b counted_at 2026-10-17T11:52:20.219Z — spec_sources.verified_at for active_params_b — when this figure was counted; a refresh that keeps it does not rewrite it confidence 0.95 — spec_sources.confidence — from the tensor headers The parameters read to produce one token. It sets the decode speed; on a dense model it equals the total. Shown to one decimal: the underlying tensor count is exact, but more digits would claim a precision the published figure does not have (SPEC §14). active	Unavailable — No quantization of this model is in the catalog yet, so there is nothing to size its weights from. (missing: model_quantizations)	Unavailable — No quantization has been sized for this model, so the weights have no measured size to work from. (missing: no quantized repos listed (GGUF/MLX) — only the native checkpoint can be sized, quantization file sizes)	Unavailable — No quantization has been sized for this model, so the weights have no measured size to work from. Nothing can be said about which machines run it until it can be sized. (missing: no quantized repos listed (GGUF/MLX) — only the native checkpoint can be sized, quantization file sizes)	ONBOARDING — No quantization has been sized for this model, so the weights have no measured size to work from. Missing: no quantized repos listed (GGUF/MLX) — only the native checkpoint can be sized, quantization file sizes.	read stale 3 d — past the 36 hour freshness window — Canary-Qwen 2.5B: the least recent read of its data from source (shape counted Oct 17, 7:52 AM ET from the tensor headers of https://huggingface.co/nvidia/canary-qwen-2.5b; no quantization size stored); never checked upstream.
Chandra OCR 2Datalab · dense · vision · 256K ctx max	4.8 (derived) DERIVEDChandra OCR 2 total parametersArithmetic on observed inputs only. No tunable assumption enters it. Σ params of language-model tensors — Cross-checked: tensor count 5.3B vs Hugging Face 5.3B (0%). 4.8 B, counted from the tensor headers of https://huggingface.co/datalab-to/chandra-ocr-2 at 2026-10-17T13:12:49.880Z total_parameters 4.8 B — https://huggingface.co/datalab-to/chandra-ocr-2 counted_at 2026-10-17T13:12:49.880Z — spec_sources.verified_at for total_params_b — when this figure was counted; a refresh that keeps it does not rewrite it confidence 0.95 — spec_sources.confidence — from the tensor headers The parameters of the language-model tensors only, counted from the tensor headers: a vision or audio tower and multi-token-prediction heads are excluded, so it can be below the whole-checkpoint count the formula's cross-check quotes. It sizes the weights only when no measured file size is stored for the chosen quantization; it is not what is read per token on a mixture-of-experts model. Shown to one decimal: the underlying tensor count is exact, but more digits would claim a precision the published figure does not have (SPEC §14). 4.2 (derived) DERIVEDChandra OCR 2 active parameters per tokenArithmetic on observed inputs only. No tunable assumption enters it. active − input embedding table — Per-token memory read used for decode bandwidth (SPEC §8.2). 4.2 B, counted from the tensor headers of https://huggingface.co/datalab-to/chandra-ocr-2 at 2026-10-17T13:12:49.880Z active_parameters_per_token 4.2 B — https://huggingface.co/datalab-to/chandra-ocr-2 counted_at 2026-10-17T13:12:49.880Z — spec_sources.verified_at for active_params_b — when this figure was counted; a refresh that keeps it does not rewrite it confidence 0.95 — spec_sources.confidence — from the tensor headers The parameters read to produce one token. It sets the decode speed; on a dense model it equals the total. Shown to one decimal: the underlying tensor count is exact, but more digits would claim a precision the published figure does not have (SPEC §14). active	Unavailable — No quantization of this model is in the catalog yet, so there is nothing to size its weights from. (missing: model_quantizations)	Unavailable — No quantization has been sized for this model, so the weights have no measured size to work from. (missing: no quantized repos listed (GGUF/MLX) — only the native checkpoint can be sized, quantization file sizes)	Unavailable — No quantization has been sized for this model, so the weights have no measured size to work from. Nothing can be said about which machines run it until it can be sized. (missing: no quantized repos listed (GGUF/MLX) — only the native checkpoint can be sized, quantization file sizes)	ONBOARDING — No quantization has been sized for this model, so the weights have no measured size to work from. Missing: no quantized repos listed (GGUF/MLX) — only the native checkpoint can be sized, quantization file sizes.	read stale 3 d — past the 36 hour freshness window — Chandra OCR 2: the least recent read of its data from source (shape counted Oct 17, 9:12 AM ET from the tensor headers of https://huggingface.co/datalab-to/chandra-ocr-2; no quantization size stored); never checked upstream.
DeepSeek-OCR 2DeepSeek · MoE · vision · 8K ctx max	3.2 (derived) DERIVEDDeepSeek-OCR 2 total parametersArithmetic on observed inputs only. No tunable assumption enters it. Σ params of language-model tensors — Cross-checked: tensor count 3.39B vs Hugging Face 3.39B (0%). 3.2 B, counted from the tensor headers of https://huggingface.co/deepseek-ai/DeepSeek-OCR-2 at 2026-10-17T13:09:03.333Z total_parameters 3.2 B — https://huggingface.co/deepseek-ai/DeepSeek-OCR-2 counted_at 2026-10-17T13:09:03.333Z — spec_sources.verified_at for total_params_b — when this figure was counted; a refresh that keeps it does not rewrite it confidence 0.95 — spec_sources.confidence — from the tensor headers The parameters of the language-model tensors only, counted from the tensor headers: a vision or audio tower and multi-token-prediction heads are excluded, so it can be below the whole-checkpoint count the formula's cross-check quotes. It sizes the weights only when no measured file size is stored for the chosen quantization; it is not what is read per token on a mixture-of-experts model. Shown to one decimal: the underlying tensor count is exact, but more digits would claim a precision the published figure does not have (SPEC §14). 0.8 (derived) DERIVEDDeepSeek-OCR 2 active parameters per tokenArithmetic on observed inputs only. No tunable assumption enters it. active − input embedding table — Per-token memory read used for decode bandwidth (SPEC §8.2). 0.8 B, counted from the tensor headers of https://huggingface.co/deepseek-ai/DeepSeek-OCR-2 at 2026-10-17T13:09:03.333Z active_parameters_per_token 0.8 B — https://huggingface.co/deepseek-ai/DeepSeek-OCR-2 counted_at 2026-10-17T13:09:03.333Z — spec_sources.verified_at for active_params_b — when this figure was counted; a refresh that keeps it does not rewrite it confidence 0.95 — spec_sources.confidence — from the tensor headers The parameters read to produce one token. It sets the decode speed; on a dense model it equals the total. Shown to one decimal: the underlying tensor count is exact, but more digits would claim a precision the published figure does not have (SPEC §14). active	Unavailable — No quantization of this model is in the catalog yet, so there is nothing to size its weights from. (missing: model_quantizations)	Unavailable — No quantization has been sized for this model, so the weights have no measured size to work from. (missing: no quantized repos listed (GGUF/MLX) — only the native checkpoint can be sized, quantization file sizes)	Unavailable — No quantization has been sized for this model, so the weights have no measured size to work from. Nothing can be said about which machines run it until it can be sized. (missing: no quantized repos listed (GGUF/MLX) — only the native checkpoint can be sized, quantization file sizes)	ONBOARDING — No quantization has been sized for this model, so the weights have no measured size to work from. Missing: no quantized repos listed (GGUF/MLX) — only the native checkpoint can be sized, quantization file sizes.	read stale 3 d — past the 36 hour freshness window — DeepSeek-OCR 2: the least recent read of its data from source (shape counted Oct 17, 9:09 AM ET from the tensor headers of https://huggingface.co/deepseek-ai/DeepSeek-OCR-2; no quantization size stored); never checked upstream.
DeepSeek-V4.1-FlashDeepSeek · MoE · reasoning, coding, vision · 1M ctx max	477.2 (derived) DERIVEDDeepSeek-V4.1-Flash total parametersArithmetic on observed inputs only. No tunable assumption enters it. Σ params of language-model tensors — UNCONFIRMED: tensor count 484.62B vs Hugging Face 763.21B (-36.5%). 477.2 B, counted from the tensor headers of https://huggingface.co/deepseek-ai/DeepSeek-V4.1-Flash at 2026-10-17T09:46:42.016Z total_parameters 477.2 B — https://huggingface.co/deepseek-ai/DeepSeek-V4.1-Flash counted_at 2026-10-17T09:46:42.016Z — spec_sources.verified_at for total_params_b — when this figure was counted; a refresh that keeps it does not rewrite it confidence 0.95 — spec_sources.confidence — from the tensor headers The parameters of the language-model tensors only, counted from the tensor headers: a vision or audio tower and multi-token-prediction heads are excluded, so it can be below the whole-checkpoint count the formula's cross-check quotes. It sizes the weights only when no measured file size is stored for the chosen quantization; it is not what is read per token on a mixture-of-experts model. Shown to one decimal: the underlying tensor count is exact, but more digits would claim a precision the published figure does not have (SPEC §14). 12.4 (derived) DERIVEDDeepSeek-V4.1-Flash active parameters per tokenArithmetic on observed inputs only. No tunable assumption enters it. active − input embedding table — Per-token memory read used for decode bandwidth (SPEC §8.2). 12.4 B, counted from the tensor headers of https://huggingface.co/deepseek-ai/DeepSeek-V4.1-Flash at 2026-10-17T09:46:42.016Z active_parameters_per_token 12.4 B — https://huggingface.co/deepseek-ai/DeepSeek-V4.1-Flash counted_at 2026-10-17T09:46:42.016Z — spec_sources.verified_at for active_params_b — when this figure was counted; a refresh that keeps it does not rewrite it confidence 0.95 — spec_sources.confidence — from the tensor headers The parameters read to produce one token. It sets the decode speed; on a dense model it equals the total. Shown to one decimal: the underlying tensor count is exact, but more digits would claim a precision the published figure does not have (SPEC §14). active	Unavailable — No quantization of this model is in the catalog yet, so there is nothing to size its weights from. (missing: model_quantizations)	Unavailable — No quantization has been sized for this model, so the weights have no measured size to work from. (missing: no quantized repos listed (GGUF/MLX) — only the native checkpoint can be sized, Plan asks for quant "4bit" but no sized quantization of DeepSeek-V4.1-Flash has that name, quantization file sizes)	Unavailable — No quantization has been sized for this model, so the weights have no measured size to work from. Nothing can be said about which machines run it until it can be sized. (missing: no quantized repos listed (GGUF/MLX) — only the native checkpoint can be sized, Plan asks for quant "4bit" but no sized quantization of DeepSeek-V4.1-Flash has that name, quantization file sizes)	ONBOARDING — No quantization has been sized for this model, so the weights have no measured size to work from. Missing: no quantized repos listed (GGUF/MLX) — only the native checkpoint can be sized, Plan asks for quant "4bit" but no sized quantization of DeepSeek-V4.1-Flash has that name, quantization file sizes.	read stale 3 d — past the 36 hour freshness window — DeepSeek-V4.1-Flash: the least recent read of its data from source (shape counted Oct 17, 5:46 AM ET from the tensor headers of https://huggingface.co/deepseek-ai/DeepSeek-V4.1-Flash; no quantization size stored); never checked upstream.
Distil-Whisper distil-large-v3.5Hugging Face (distil-whisper) · dense · speech to text	0.8 (derived) DERIVEDDistil-Whisper distil-large-v3.5 total parametersArithmetic on observed inputs only. No tunable assumption enters it. Σ params of language-model tensors — Cross-checked: tensor count 0.76B vs Hugging Face 0.76B (0%). 0.8 B, counted from the tensor headers of https://huggingface.co/distil-whisper/distil-large-v3.5 at 2026-10-17T11:55:27.547Z total_parameters 0.8 B — https://huggingface.co/distil-whisper/distil-large-v3.5 counted_at 2026-10-17T11:55:27.547Z — spec_sources.verified_at for total_params_b — when this figure was counted; a refresh that keeps it does not rewrite it confidence 0.95 — spec_sources.confidence — from the tensor headers The parameters of the language-model tensors only, counted from the tensor headers: a vision or audio tower and multi-token-prediction heads are excluded, so it can be below the whole-checkpoint count the formula's cross-check quotes. It sizes the weights only when no measured file size is stored for the chosen quantization; it is not what is read per token on a mixture-of-experts model. Shown to one decimal: the underlying tensor count is exact, but more digits would claim a precision the published figure does not have (SPEC §14). 0.7 (derived) DERIVEDDistil-Whisper distil-large-v3.5 active parameters per tokenArithmetic on observed inputs only. No tunable assumption enters it. active − input embedding table — Per-token memory read used for decode bandwidth (SPEC §8.2). 0.7 B, counted from the tensor headers of https://huggingface.co/distil-whisper/distil-large-v3.5 at 2026-10-17T11:55:27.547Z active_parameters_per_token 0.7 B — https://huggingface.co/distil-whisper/distil-large-v3.5 counted_at 2026-10-17T11:55:27.547Z — spec_sources.verified_at for active_params_b — when this figure was counted; a refresh that keeps it does not rewrite it confidence 0.95 — spec_sources.confidence — from the tensor headers The parameters read to produce one token. It sets the decode speed; on a dense model it equals the total. Shown to one decimal: the underlying tensor count is exact, but more digits would claim a precision the published figure does not have (SPEC §14). active	Unavailable — No quantization of this model is in the catalog yet, so there is nothing to size its weights from. (missing: model_quantizations)	Unavailable — No quantization has been sized for this model, so the weights have no measured size to work from. (missing: no quantized repos listed (GGUF/MLX) — only the native checkpoint can be sized, quantization file sizes)	Unavailable — No quantization has been sized for this model, so the weights have no measured size to work from. Nothing can be said about which machines run it until it can be sized. (missing: no quantized repos listed (GGUF/MLX) — only the native checkpoint can be sized, quantization file sizes)	ONBOARDING — No quantization has been sized for this model, so the weights have no measured size to work from. Missing: no quantized repos listed (GGUF/MLX) — only the native checkpoint can be sized, quantization file sizes.	read stale 3 d — past the 36 hour freshness window — Distil-Whisper distil-large-v3.5: the least recent read of its data from source (shape counted Oct 17, 7:55 AM ET from the tensor headers of https://huggingface.co/distil-whisper/distil-large-v3.5; no quantization size stored); never checked upstream.
dots.ocrrednote (dots-studio) · dense · vision · 128K ctx max	1.8 (derived) DERIVEDdots.ocr total parametersArithmetic on observed inputs only. No tunable assumption enters it. Σ params of language-model tensors — Cross-checked: tensor count 3.04B vs Hugging Face 3.04B (0%). 1.8 B, counted from the tensor headers of https://huggingface.co/dots-studio/dots.ocr at 2026-10-17T13:11:45.777Z total_parameters 1.8 B — https://huggingface.co/dots-studio/dots.ocr counted_at 2026-10-17T13:11:45.777Z — spec_sources.verified_at for total_params_b — when this figure was counted; a refresh that keeps it does not rewrite it confidence 0.95 — spec_sources.confidence — from the tensor headers The parameters of the language-model tensors only, counted from the tensor headers: a vision or audio tower and multi-token-prediction heads are excluded, so it can be below the whole-checkpoint count the formula's cross-check quotes. It sizes the weights only when no measured file size is stored for the chosen quantization; it is not what is read per token on a mixture-of-experts model. Shown to one decimal: the underlying tensor count is exact, but more digits would claim a precision the published figure does not have (SPEC §14). 1.5 (derived) DERIVEDdots.ocr active parameters per tokenArithmetic on observed inputs only. No tunable assumption enters it. active − input embedding table — Per-token memory read used for decode bandwidth (SPEC §8.2). 1.5 B, counted from the tensor headers of https://huggingface.co/dots-studio/dots.ocr at 2026-10-17T13:11:45.777Z active_parameters_per_token 1.5 B — https://huggingface.co/dots-studio/dots.ocr counted_at 2026-10-17T13:11:45.777Z — spec_sources.verified_at for active_params_b — when this figure was counted; a refresh that keeps it does not rewrite it confidence 0.95 — spec_sources.confidence — from the tensor headers The parameters read to produce one token. It sets the decode speed; on a dense model it equals the total. Shown to one decimal: the underlying tensor count is exact, but more digits would claim a precision the published figure does not have (SPEC §14). active	Unavailable — No quantization of this model is in the catalog yet, so there is nothing to size its weights from. (missing: model_quantizations)	Unavailable — No quantization has been sized for this model, so the weights have no measured size to work from. (missing: no quantized repos listed (GGUF/MLX) — only the native checkpoint can be sized, quantization file sizes)	Unavailable — No quantization has been sized for this model, so the weights have no measured size to work from. Nothing can be said about which machines run it until it can be sized. (missing: no quantized repos listed (GGUF/MLX) — only the native checkpoint can be sized, quantization file sizes)	ONBOARDING — No quantization has been sized for this model, so the weights have no measured size to work from. Missing: no quantized repos listed (GGUF/MLX) — only the native checkpoint can be sized, quantization file sizes.	read stale 3 d — past the 36 hour freshness window — dots.ocr: the least recent read of its data from source (shape counted Oct 17, 9:11 AM ET from the tensor headers of https://huggingface.co/dots-studio/dots.ocr; no quantization size stored); never checked upstream.
Gemma 4 26B-A4B ITGoogle · MoE · vision, reasoning · 256K ctx max	Unavailable — The total parameters of Gemma 4 26B-A4B IT has not been read from its checkpoint yet, so nothing that depends on it can be computed. (missing: total_parameters)	Unavailable — No quantization of this model is in the catalog yet, so there is nothing to size its weights from. (missing: model_quantizations)	Unavailable — The model's shape has not been onboarded, so its weights and KV cache cannot be computed. (missing: model shape (tensor inventory and config formula both failed), bits per weight for bartowski/google_gemma-4-26B-A4B-it-GGUF:IQ2_XXS (model shape unknown), bits per weight for bartowski/google_gemma-4-26B-A4B-it-GGUF:IQ2_XS (model shape unknown), bits per weight for bartowski/google_gemma-4-26B-A4B-it-GGUF:IQ2_S (model shape unknown), bits per weight for bartowski/google_gemma-4-26B-A4B-it-GGUF:IQ2_M (model shape unknown), bits per weight for bartowski/google_gemma-4-26B-A4B-it-GGUF:Q2_K (model shape unknown), bits per weight for bartowski/google_gemma-4-26B-A4B-it-GGUF:Q2_K_L (model shape unknown), bits per weight for bartowski/google_gemma-4-26B-A4B-it-GGUF:IQ3_XXS (model shape unknown), bits per weight for bartowski/google_gemma-4-26B-A4B-it-GGUF:IQ3_XS (model shape unknown), bits per weight for bartowski/google_gemma-4-26B-A4B-it-GGUF:Q3_K_S (model shape unknown), bits per weight for bartowski/google_gemma-4-26B-A4B-it-GGUF:Q3_K_M (model shape unknown), bits per weight for bartowski/google_gemma-4-26B-A4B-it-GGUF:Q3_K_L (model shape unknown), bits per weight for bartowski/google_gemma-4-26B-A4B-it-GGUF:IQ3_M (model shape unknown), bits per weight for bartowski/google_gemma-4-26B-A4B-it-GGUF:Q3_K_XL (model shape unknown), bits per weight for bartowski/google_gemma-4-26B-A4B-it-GGUF:IQ4_XS (model shape unknown), bits per weight for bartowski/google_gemma-4-26B-A4B-it-GGUF:IQ4_NL (model shape unknown), bits per weight for bartowski/google_gemma-4-26B-A4B-it-GGUF:Q4_0 (model shape unknown), bits per weight for bartowski/google_gemma-4-26B-A4B-it-GGUF:Q4_K_S (model shape unknown), bits per weight for bartowski/google_gemma-4-26B-A4B-it-GGUF:Q4_1 (model shape unknown), bits per weight for bartowski/google_gemma-4-26B-A4B-it-GGUF:Q4_K_M (model shape unknown), bits per weight for bartowski/google_gemma-4-26B-A4B-it-GGUF:Q4_K_L (model shape unknown), bits per weight for bartowski/google_gemma-4-26B-A4B-it-GGUF:Q5_K_S (model shape unknown), bits per weight for bartowski/google_gemma-4-26B-A4B-it-GGUF:Q5_K_M (model shape unknown), bits per weight for bartowski/google_gemma-4-26B-A4B-it-GGUF:Q5_K_L (model shape unknown), bits per weight for bartowski/google_gemma-4-26B-A4B-it-GGUF:Q6_K (model shape unknown), bits per weight for bartowski/google_gemma-4-26B-A4B-it-GGUF:Q6_K_L (model shape unknown), bits per weight for bartowski/google_gemma-4-26B-A4B-it-GGUF:Q8_0 (model shape unknown), bits per weight for bartowski/google_gemma-4-26B-A4B-it-GGUF:BF16 (model shape unknown), bits per weight for ggml-org/gemma-4-26B-A4B-it-GGUF:Q4_0 (model shape unknown), bits per weight for ggml-org/gemma-4-26B-A4B-it-GGUF:Q8_0 (model shape unknown), bits per weight for ggml-org/gemma-4-26B-A4B-it-GGUF:BF16 (model shape unknown), bits per weight for lmstudio-community/gemma-4-26B-A4B-it-GGUF:Q4_K_M (model shape unknown), bits per weight for lmstudio-community/gemma-4-26B-A4B-it-GGUF:Q6_K (model shape unknown), bits per weight for lmstudio-community/gemma-4-26B-A4B-it-GGUF:Q8_0 (model shape unknown), bits per weight for lmstudio-community/gemma-4-26B-A4B-it-MLX-4bit:4bit (model shape unknown), bits per weight for lmstudio-community/gemma-4-26B-A4B-it-MLX-5bit:5bit (model shape unknown), bits per weight for lmstudio-community/gemma-4-26B-A4B-it-MLX-6bit:6bit (model shape unknown), bits per weight for lmstudio-community/gemma-4-26B-A4B-it-MLX-8bit:8bit (model shape unknown), bits per weight for mlx-community/gemma-4-26B-A4B-it-OptiQ-4bit:4bit (model shape unknown), bits per weight for mlx-community/gemma-4-26B-A4B-it-heretic-msq-2.6bit:2.6bit (model shape unknown), bits per weight for mlx-community/gemma-4-26b-a4b-it-4bit:4bit (model shape unknown), bits per weight for mlx-community/gemma-4-26b-a4b-it-6bit:6bit (model shape unknown), bits per weight for mlx-community/gemma-4-26b-a4b-it-8bit:8bit (model shape unknown), bits per weight for mlx-community/gemma-4-26b-a4b-it-mxfp8:mxfp8 (model shape unknown), bits per weight for mlx-community/gemma-4-26b-a4b-it-nvfp4:native (model shape unknown), bits per weight for unsloth/gemma-4-26B-A4B-it-GGUF:F16 (model shape unknown), bits per weight for unsloth/gemma-4-26B-A4B-it-GGUF:UD-IQ2_XXS (model shape unknown), bits per weight for unsloth/gemma-4-26B-A4B-it-GGUF:UD-IQ2_M (model shape unknown), bits per weight for unsloth/gemma-4-26B-A4B-it-GGUF:UD-Q2_K_XL (model shape unknown), bits per weight for unsloth/gemma-4-26B-A4B-it-GGUF:UD-IQ3_S (model shape unknown), bits per weight for unsloth/gemma-4-26B-A4B-it-GGUF:UD-IQ3_XXS (model shape unknown), bits per weight for unsloth/gemma-4-26B-A4B-it-GGUF:UD-Q3_K_M (model shape unknown), bits per weight for unsloth/gemma-4-26B-A4B-it-GGUF:UD-Q3_K_XL (model shape unknown), bits per weight for unsloth/gemma-4-26B-A4B-it-GGUF:UD-IQ4_XS (model shape unknown), bits per weight for unsloth/gemma-4-26B-A4B-it-GGUF:UD-IQ4_NL (model shape unknown), bits per weight for unsloth/gemma-4-26B-A4B-it-GGUF:UD-Q4_K_S (model shape unknown), bits per weight for unsloth/gemma-4-26B-A4B-it-GGUF:MXFP4_MOE (model shape unknown), bits per weight for unsloth/gemma-4-26B-A4B-it-GGUF:UD-Q4_K_M (model shape unknown), bits per weight for unsloth/gemma-4-26B-A4B-it-GGUF:UD-Q4_K_XL (model shape unknown), bits per weight for unsloth/gemma-4-26B-A4B-it-GGUF:UD-Q5_K_S (model shape unknown), bits per weight for unsloth/gemma-4-26B-A4B-it-GGUF:UD-Q5_K_M (model shape unknown), bits per weight for unsloth/gemma-4-26B-A4B-it-GGUF:UD-Q5_K_XL (model shape unknown), bits per weight for unsloth/gemma-4-26B-A4B-it-GGUF:UD-Q6_K (model shape unknown), bits per weight for unsloth/gemma-4-26B-A4B-it-GGUF:UD-Q6_K_XL (model shape unknown), bits per weight for unsloth/gemma-4-26B-A4B-it-GGUF:Q8_0 (model shape unknown), bits per weight for unsloth/gemma-4-26B-A4B-it-GGUF:UD-Q8_K_XL (model shape unknown), bits per weight for unsloth/gemma-4-26B-A4B-it-GGUF:BF16 (model shape unknown), bits per weight for unsloth/gemma-4-26B-A4B-it-NVFP4:native (model shape unknown), bits per weight for unsloth/gemma-4-26b-a4b-it-MLX-8bit:8bit (model shape unknown), bits per weight for unsloth/gemma-4-26b-a4b-it-UD-MLX-3bit:3bit (model shape unknown), bits per weight for unsloth/gemma-4-26b-a4b-it-UD-MLX-4bit:4bit (model shape unknown), total_params_b, active_params_b, num_layers, attention_params)	Unavailable — The model's shape has not been onboarded, so its weights and KV cache cannot be computed. Nothing can be said about which machines run it until it can be sized. (missing: model shape (tensor inventory and config formula both failed), bits per weight for bartowski/google_gemma-4-26B-A4B-it-GGUF:IQ2_XXS (model shape unknown), bits per weight for bartowski/google_gemma-4-26B-A4B-it-GGUF:IQ2_XS (model shape unknown), bits per weight for bartowski/google_gemma-4-26B-A4B-it-GGUF:IQ2_S (model shape unknown), bits per weight for bartowski/google_gemma-4-26B-A4B-it-GGUF:IQ2_M (model shape unknown), bits per weight for bartowski/google_gemma-4-26B-A4B-it-GGUF:Q2_K (model shape unknown), bits per weight for bartowski/google_gemma-4-26B-A4B-it-GGUF:Q2_K_L (model shape unknown), bits per weight for bartowski/google_gemma-4-26B-A4B-it-GGUF:IQ3_XXS (model shape unknown), bits per weight for bartowski/google_gemma-4-26B-A4B-it-GGUF:IQ3_XS (model shape unknown), bits per weight for bartowski/google_gemma-4-26B-A4B-it-GGUF:Q3_K_S (model shape unknown), bits per weight for bartowski/google_gemma-4-26B-A4B-it-GGUF:Q3_K_M (model shape unknown), bits per weight for bartowski/google_gemma-4-26B-A4B-it-GGUF:Q3_K_L (model shape unknown), bits per weight for bartowski/google_gemma-4-26B-A4B-it-GGUF:IQ3_M (model shape unknown), bits per weight for bartowski/google_gemma-4-26B-A4B-it-GGUF:Q3_K_XL (model shape unknown), bits per weight for bartowski/google_gemma-4-26B-A4B-it-GGUF:IQ4_XS (model shape unknown), bits per weight for bartowski/google_gemma-4-26B-A4B-it-GGUF:IQ4_NL (model shape unknown), bits per weight for bartowski/google_gemma-4-26B-A4B-it-GGUF:Q4_0 (model shape unknown), bits per weight for bartowski/google_gemma-4-26B-A4B-it-GGUF:Q4_K_S (model shape unknown), bits per weight for bartowski/google_gemma-4-26B-A4B-it-GGUF:Q4_1 (model shape unknown), bits per weight for bartowski/google_gemma-4-26B-A4B-it-GGUF:Q4_K_M (model shape unknown), bits per weight for bartowski/google_gemma-4-26B-A4B-it-GGUF:Q4_K_L (model shape unknown), bits per weight for bartowski/google_gemma-4-26B-A4B-it-GGUF:Q5_K_S (model shape unknown), bits per weight for bartowski/google_gemma-4-26B-A4B-it-GGUF:Q5_K_M (model shape unknown), bits per weight for bartowski/google_gemma-4-26B-A4B-it-GGUF:Q5_K_L (model shape unknown), bits per weight for bartowski/google_gemma-4-26B-A4B-it-GGUF:Q6_K (model shape unknown), bits per weight for bartowski/google_gemma-4-26B-A4B-it-GGUF:Q6_K_L (model shape unknown), bits per weight for bartowski/google_gemma-4-26B-A4B-it-GGUF:Q8_0 (model shape unknown), bits per weight for bartowski/google_gemma-4-26B-A4B-it-GGUF:BF16 (model shape unknown), bits per weight for ggml-org/gemma-4-26B-A4B-it-GGUF:Q4_0 (model shape unknown), bits per weight for ggml-org/gemma-4-26B-A4B-it-GGUF:Q8_0 (model shape unknown), bits per weight for ggml-org/gemma-4-26B-A4B-it-GGUF:BF16 (model shape unknown), bits per weight for lmstudio-community/gemma-4-26B-A4B-it-GGUF:Q4_K_M (model shape unknown), bits per weight for lmstudio-community/gemma-4-26B-A4B-it-GGUF:Q6_K (model shape unknown), bits per weight for lmstudio-community/gemma-4-26B-A4B-it-GGUF:Q8_0 (model shape unknown), bits per weight for lmstudio-community/gemma-4-26B-A4B-it-MLX-4bit:4bit (model shape unknown), bits per weight for lmstudio-community/gemma-4-26B-A4B-it-MLX-5bit:5bit (model shape unknown), bits per weight for lmstudio-community/gemma-4-26B-A4B-it-MLX-6bit:6bit (model shape unknown), bits per weight for lmstudio-community/gemma-4-26B-A4B-it-MLX-8bit:8bit (model shape unknown), bits per weight for mlx-community/gemma-4-26B-A4B-it-OptiQ-4bit:4bit (model shape unknown), bits per weight for mlx-community/gemma-4-26B-A4B-it-heretic-msq-2.6bit:2.6bit (model shape unknown), bits per weight for mlx-community/gemma-4-26b-a4b-it-4bit:4bit (model shape unknown), bits per weight for mlx-community/gemma-4-26b-a4b-it-6bit:6bit (model shape unknown), bits per weight for mlx-community/gemma-4-26b-a4b-it-8bit:8bit (model shape unknown), bits per weight for mlx-community/gemma-4-26b-a4b-it-mxfp8:mxfp8 (model shape unknown), bits per weight for mlx-community/gemma-4-26b-a4b-it-nvfp4:native (model shape unknown), bits per weight for unsloth/gemma-4-26B-A4B-it-GGUF:F16 (model shape unknown), bits per weight for unsloth/gemma-4-26B-A4B-it-GGUF:UD-IQ2_XXS (model shape unknown), bits per weight for unsloth/gemma-4-26B-A4B-it-GGUF:UD-IQ2_M (model shape unknown), bits per weight for unsloth/gemma-4-26B-A4B-it-GGUF:UD-Q2_K_XL (model shape unknown), bits per weight for unsloth/gemma-4-26B-A4B-it-GGUF:UD-IQ3_S (model shape unknown), bits per weight for unsloth/gemma-4-26B-A4B-it-GGUF:UD-IQ3_XXS (model shape unknown), bits per weight for unsloth/gemma-4-26B-A4B-it-GGUF:UD-Q3_K_M (model shape unknown), bits per weight for unsloth/gemma-4-26B-A4B-it-GGUF:UD-Q3_K_XL (model shape unknown), bits per weight for unsloth/gemma-4-26B-A4B-it-GGUF:UD-IQ4_XS (model shape unknown), bits per weight for unsloth/gemma-4-26B-A4B-it-GGUF:UD-IQ4_NL (model shape unknown), bits per weight for unsloth/gemma-4-26B-A4B-it-GGUF:UD-Q4_K_S (model shape unknown), bits per weight for unsloth/gemma-4-26B-A4B-it-GGUF:MXFP4_MOE (model shape unknown), bits per weight for unsloth/gemma-4-26B-A4B-it-GGUF:UD-Q4_K_M (model shape unknown), bits per weight for unsloth/gemma-4-26B-A4B-it-GGUF:UD-Q4_K_XL (model shape unknown), bits per weight for unsloth/gemma-4-26B-A4B-it-GGUF:UD-Q5_K_S (model shape unknown), bits per weight for unsloth/gemma-4-26B-A4B-it-GGUF:UD-Q5_K_M (model shape unknown), bits per weight for unsloth/gemma-4-26B-A4B-it-GGUF:UD-Q5_K_XL (model shape unknown), bits per weight for unsloth/gemma-4-26B-A4B-it-GGUF:UD-Q6_K (model shape unknown), bits per weight for unsloth/gemma-4-26B-A4B-it-GGUF:UD-Q6_K_XL (model shape unknown), bits per weight for unsloth/gemma-4-26B-A4B-it-GGUF:Q8_0 (model shape unknown), bits per weight for unsloth/gemma-4-26B-A4B-it-GGUF:UD-Q8_K_XL (model shape unknown), bits per weight for unsloth/gemma-4-26B-A4B-it-GGUF:BF16 (model shape unknown), bits per weight for unsloth/gemma-4-26B-A4B-it-NVFP4:native (model shape unknown), bits per weight for unsloth/gemma-4-26b-a4b-it-MLX-8bit:8bit (model shape unknown), bits per weight for unsloth/gemma-4-26b-a4b-it-UD-MLX-3bit:3bit (model shape unknown), bits per weight for unsloth/gemma-4-26b-a4b-it-UD-MLX-4bit:4bit (model shape unknown), total_params_b, active_params_b, num_layers, attention_params)	ONBOARDING — The model's shape has not been onboarded, so its weights and KV cache cannot be computed. Missing: model shape (tensor inventory and config formula both failed), bits per weight for bartowski/google_gemma-4-26B-A4B-it-GGUF:IQ2_XXS (model shape unknown), bits per weight for bartowski/google_gemma-4-26B-A4B-it-GGUF:IQ2_XS (model shape unknown), bits per weight for bartowski/google_gemma-4-26B-A4B-it-GGUF:IQ2_S (model shape unknown), bits per weight for bartowski/google_gemma-4-26B-A4B-it-GGUF:IQ2_M (model shape unknown), bits per weight for bartowski/google_gemma-4-26B-A4B-it-GGUF:Q2_K (model shape unknown), bits per weight for bartowski/google_gemma-4-26B-A4B-it-GGUF:Q2_K_L (model shape unknown), bits per weight for bartowski/google_gemma-4-26B-A4B-it-GGUF:IQ3_XXS (model shape unknown), bits per weight for bartowski/google_gemma-4-26B-A4B-it-GGUF:IQ3_XS (model shape unknown), bits per weight for bartowski/google_gemma-4-26B-A4B-it-GGUF:Q3_K_S (model shape unknown), bits per weight for bartowski/google_gemma-4-26B-A4B-it-GGUF:Q3_K_M (model shape unknown), bits per weight for bartowski/google_gemma-4-26B-A4B-it-GGUF:Q3_K_L (model shape unknown), bits per weight for bartowski/google_gemma-4-26B-A4B-it-GGUF:IQ3_M (model shape unknown), bits per weight for bartowski/google_gemma-4-26B-A4B-it-GGUF:Q3_K_XL (model shape unknown), bits per weight for bartowski/google_gemma-4-26B-A4B-it-GGUF:IQ4_XS (model shape unknown), bits per weight for bartowski/google_gemma-4-26B-A4B-it-GGUF:IQ4_NL (model shape unknown), bits per weight for bartowski/google_gemma-4-26B-A4B-it-GGUF:Q4_0 (model shape unknown), bits per weight for bartowski/google_gemma-4-26B-A4B-it-GGUF:Q4_K_S (model shape unknown), bits per weight for bartowski/google_gemma-4-26B-A4B-it-GGUF:Q4_1 (model shape unknown), bits per weight for bartowski/google_gemma-4-26B-A4B-it-GGUF:Q4_K_M (model shape unknown), bits per weight for bartowski/google_gemma-4-26B-A4B-it-GGUF:Q4_K_L (model shape unknown), bits per weight for bartowski/google_gemma-4-26B-A4B-it-GGUF:Q5_K_S (model shape unknown), bits per weight for bartowski/google_gemma-4-26B-A4B-it-GGUF:Q5_K_M (model shape unknown), bits per weight for bartowski/google_gemma-4-26B-A4B-it-GGUF:Q5_K_L (model shape unknown), bits per weight for bartowski/google_gemma-4-26B-A4B-it-GGUF:Q6_K (model shape unknown), bits per weight for bartowski/google_gemma-4-26B-A4B-it-GGUF:Q6_K_L (model shape unknown), bits per weight for bartowski/google_gemma-4-26B-A4B-it-GGUF:Q8_0 (model shape unknown), bits per weight for bartowski/google_gemma-4-26B-A4B-it-GGUF:BF16 (model shape unknown), bits per weight for ggml-org/gemma-4-26B-A4B-it-GGUF:Q4_0 (model shape unknown), bits per weight for ggml-org/gemma-4-26B-A4B-it-GGUF:Q8_0 (model shape unknown), bits per weight for ggml-org/gemma-4-26B-A4B-it-GGUF:BF16 (model shape unknown), bits per weight for lmstudio-community/gemma-4-26B-A4B-it-GGUF:Q4_K_M (model shape unknown), bits per weight for lmstudio-community/gemma-4-26B-A4B-it-GGUF:Q6_K (model shape unknown), bits per weight for lmstudio-community/gemma-4-26B-A4B-it-GGUF:Q8_0 (model shape unknown), bits per weight for lmstudio-community/gemma-4-26B-A4B-it-MLX-4bit:4bit (model shape unknown), bits per weight for lmstudio-community/gemma-4-26B-A4B-it-MLX-5bit:5bit (model shape unknown), bits per weight for lmstudio-community/gemma-4-26B-A4B-it-MLX-6bit:6bit (model shape unknown), bits per weight for lmstudio-community/gemma-4-26B-A4B-it-MLX-8bit:8bit (model shape unknown), bits per weight for mlx-community/gemma-4-26B-A4B-it-OptiQ-4bit:4bit (model shape unknown), bits per weight for mlx-community/gemma-4-26B-A4B-it-heretic-msq-2.6bit:2.6bit (model shape unknown), bits per weight for mlx-community/gemma-4-26b-a4b-it-4bit:4bit (model shape unknown), bits per weight for mlx-community/gemma-4-26b-a4b-it-6bit:6bit (model shape unknown), bits per weight for mlx-community/gemma-4-26b-a4b-it-8bit:8bit (model shape unknown), bits per weight for mlx-community/gemma-4-26b-a4b-it-mxfp8:mxfp8 (model shape unknown), bits per weight for mlx-community/gemma-4-26b-a4b-it-nvfp4:native (model shape unknown), bits per weight for unsloth/gemma-4-26B-A4B-it-GGUF:F16 (model shape unknown), bits per weight for unsloth/gemma-4-26B-A4B-it-GGUF:UD-IQ2_XXS (model shape unknown), bits per weight for unsloth/gemma-4-26B-A4B-it-GGUF:UD-IQ2_M (model shape unknown), bits per weight for unsloth/gemma-4-26B-A4B-it-GGUF:UD-Q2_K_XL (model shape unknown), bits per weight for unsloth/gemma-4-26B-A4B-it-GGUF:UD-IQ3_S (model shape unknown), bits per weight for unsloth/gemma-4-26B-A4B-it-GGUF:UD-IQ3_XXS (model shape unknown), bits per weight for unsloth/gemma-4-26B-A4B-it-GGUF:UD-Q3_K_M (model shape unknown), bits per weight for unsloth/gemma-4-26B-A4B-it-GGUF:UD-Q3_K_XL (model shape unknown), bits per weight for unsloth/gemma-4-26B-A4B-it-GGUF:UD-IQ4_XS (model shape unknown), bits per weight for unsloth/gemma-4-26B-A4B-it-GGUF:UD-IQ4_NL (model shape unknown), bits per weight for unsloth/gemma-4-26B-A4B-it-GGUF:UD-Q4_K_S (model shape unknown), bits per weight for unsloth/gemma-4-26B-A4B-it-GGUF:MXFP4_MOE (model shape unknown), bits per weight for unsloth/gemma-4-26B-A4B-it-GGUF:UD-Q4_K_M (model shape unknown), bits per weight for unsloth/gemma-4-26B-A4B-it-GGUF:UD-Q4_K_XL (model shape unknown), bits per weight for unsloth/gemma-4-26B-A4B-it-GGUF:UD-Q5_K_S (model shape unknown), bits per weight for unsloth/gemma-4-26B-A4B-it-GGUF:UD-Q5_K_M (model shape unknown), bits per weight for unsloth/gemma-4-26B-A4B-it-GGUF:UD-Q5_K_XL (model shape unknown), bits per weight for unsloth/gemma-4-26B-A4B-it-GGUF:UD-Q6_K (model shape unknown), bits per weight for unsloth/gemma-4-26B-A4B-it-GGUF:UD-Q6_K_XL (model shape unknown), bits per weight for unsloth/gemma-4-26B-A4B-it-GGUF:Q8_0 (model shape unknown), bits per weight for unsloth/gemma-4-26B-A4B-it-GGUF:UD-Q8_K_XL (model shape unknown), bits per weight for unsloth/gemma-4-26B-A4B-it-GGUF:BF16 (model shape unknown), bits per weight for unsloth/gemma-4-26B-A4B-it-NVFP4:native (model shape unknown), bits per weight for unsloth/gemma-4-26b-a4b-it-MLX-8bit:8bit (model shape unknown), bits per weight for unsloth/gemma-4-26b-a4b-it-UD-MLX-3bit:3bit (model shape unknown), bits per weight for unsloth/gemma-4-26b-a4b-it-UD-MLX-4bit:4bit (model shape unknown), total_params_b, active_params_b, num_layers, attention_params.	Unavailable — Gemma 4 26B-A4B IT: No shape count (spec_sources.total_params_b) of google/gemma-4-26B-A4B-it and no stored quantization size (a model_quantizations row with a repository and a read time) is on record, so how current this model's data is cannot be dated. Other fields — licence, maximum context — may still have been read from its config, and a quantization listing may have been read without yielding a size; the onboarding state says what is owed. (missing: spec_sources.total_params_b, model_quantizations.verified_at)
Granite Speech 4.1 2BIBM · dense · speech to text · 4K ctx max	2.3 (derived) DERIVEDGranite Speech 4.1 2B total parametersArithmetic on observed inputs only. No tunable assumption enters it. Σ params of language-model tensors — Cross-checked: tensor count 2.31B vs Hugging Face 2.31B (6.92e-7%). 2.3 B, counted from the tensor headers of https://huggingface.co/ibm-granite/granite-speech-4.1-2b at 2026-10-17T11:52:56.967Z total_parameters 2.3 B — https://huggingface.co/ibm-granite/granite-speech-4.1-2b counted_at 2026-10-17T11:52:56.967Z — spec_sources.verified_at for total_params_b — when this figure was counted; a refresh that keeps it does not rewrite it confidence 0.95 — spec_sources.confidence — from the tensor headers The parameters of the language-model tensors only, counted from the tensor headers: a vision or audio tower and multi-token-prediction heads are excluded, so it can be below the whole-checkpoint count the formula's cross-check quotes. It sizes the weights only when no measured file size is stored for the chosen quantization; it is not what is read per token on a mixture-of-experts model. Shown to one decimal: the underlying tensor count is exact, but more digits would claim a precision the published figure does not have (SPEC §14). 2.1 (derived) DERIVEDGranite Speech 4.1 2B active parameters per tokenArithmetic on observed inputs only. No tunable assumption enters it. active − input embedding table — Per-token memory read used for decode bandwidth (SPEC §8.2). 2.1 B, counted from the tensor headers of https://huggingface.co/ibm-granite/granite-speech-4.1-2b at 2026-10-17T11:52:56.967Z active_parameters_per_token 2.1 B — https://huggingface.co/ibm-granite/granite-speech-4.1-2b counted_at 2026-10-17T11:52:56.967Z — spec_sources.verified_at for active_params_b — when this figure was counted; a refresh that keeps it does not rewrite it confidence 0.95 — spec_sources.confidence — from the tensor headers The parameters read to produce one token. It sets the decode speed; on a dense model it equals the total. Shown to one decimal: the underlying tensor count is exact, but more digits would claim a precision the published figure does not have (SPEC §14). active	Unavailable — No quantization of this model is in the catalog yet, so there is nothing to size its weights from. (missing: model_quantizations)	Unavailable — No quantization has been sized for this model, so the weights have no measured size to work from. (missing: no quantized repos listed (GGUF/MLX) — only the native checkpoint can be sized, quantization file sizes)	Unavailable — No quantization has been sized for this model, so the weights have no measured size to work from. Nothing can be said about which machines run it until it can be sized. (missing: no quantized repos listed (GGUF/MLX) — only the native checkpoint can be sized, quantization file sizes)	ONBOARDING — No quantization has been sized for this model, so the weights have no measured size to work from. Missing: no quantized repos listed (GGUF/MLX) — only the native checkpoint can be sized, quantization file sizes.	read stale 3 d — past the 36 hour freshness window — Granite Speech 4.1 2B: the least recent read of its data from source (shape counted Oct 17, 7:52 AM ET from the tensor headers of https://huggingface.co/ibm-granite/granite-speech-4.1-2b; no quantization size stored); never checked upstream.
PaddleOCR-VL (0.9B)Baidu PaddlePaddle · dense · vision · 128K ctx max	0.5 (derived) DERIVEDPaddleOCR-VL (0.9B) total parametersArithmetic on observed inputs only. No tunable assumption enters it. Σ params of language-model tensors — Cross-checked: tensor count 0.96B vs Hugging Face 0.96B (0%). 0.5 B, counted from the tensor headers of https://huggingface.co/PaddlePaddle/PaddleOCR-VL at 2026-10-17T13:11:15.283Z total_parameters 0.5 B — https://huggingface.co/PaddlePaddle/PaddleOCR-VL counted_at 2026-10-17T13:11:15.283Z — spec_sources.verified_at for total_params_b — when this figure was counted; a refresh that keeps it does not rewrite it confidence 0.95 — spec_sources.confidence — from the tensor headers The parameters of the language-model tensors only, counted from the tensor headers: a vision or audio tower and multi-token-prediction heads are excluded, so it can be below the whole-checkpoint count the formula's cross-check quotes. It sizes the weights only when no measured file size is stored for the chosen quantization; it is not what is read per token on a mixture-of-experts model. Shown to one decimal: the underlying tensor count is exact, but more digits would claim a precision the published figure does not have (SPEC §14). 0.4 (derived) DERIVEDPaddleOCR-VL (0.9B) active parameters per tokenArithmetic on observed inputs only. No tunable assumption enters it. active − input embedding table — Per-token memory read used for decode bandwidth (SPEC §8.2). 0.4 B, counted from the tensor headers of https://huggingface.co/PaddlePaddle/PaddleOCR-VL at 2026-10-17T13:11:15.283Z active_parameters_per_token 0.4 B — https://huggingface.co/PaddlePaddle/PaddleOCR-VL counted_at 2026-10-17T13:11:15.283Z — spec_sources.verified_at for active_params_b — when this figure was counted; a refresh that keeps it does not rewrite it confidence 0.95 — spec_sources.confidence — from the tensor headers The parameters read to produce one token. It sets the decode speed; on a dense model it equals the total. Shown to one decimal: the underlying tensor count is exact, but more digits would claim a precision the published figure does not have (SPEC §14). active	Unavailable — No quantization of this model is in the catalog yet, so there is nothing to size its weights from. (missing: model_quantizations)	Unavailable — No quantization has been sized for this model, so the weights have no measured size to work from. (missing: no quantized repos listed (GGUF/MLX) — only the native checkpoint can be sized, quantization file sizes)	Unavailable — No quantization has been sized for this model, so the weights have no measured size to work from. Nothing can be said about which machines run it until it can be sized. (missing: no quantized repos listed (GGUF/MLX) — only the native checkpoint can be sized, quantization file sizes)	ONBOARDING — No quantization has been sized for this model, so the weights have no measured size to work from. Missing: no quantized repos listed (GGUF/MLX) — only the native checkpoint can be sized, quantization file sizes.	read stale 3 d — past the 36 hour freshness window — PaddleOCR-VL (0.9B): the least recent read of its data from source (shape counted Oct 17, 9:11 AM ET from the tensor headers of https://huggingface.co/PaddlePaddle/PaddleOCR-VL; no quantization size stored); never checked upstream.
Parakeet TDT 0.6B v3NVIDIA · dense · speech to text	0.6 (derived) DERIVEDParakeet TDT 0.6B v3 total parametersArithmetic on observed inputs only. No tunable assumption enters it. Σ params of language-model tensors — Cross-checked: tensor count 0.63B vs Hugging Face 0.63B (3.83e-6%). 0.6 B, counted from the tensor headers of https://huggingface.co/nvidia/parakeet-tdt-0.6b-v3 at 2026-10-17T11:51:44.474Z total_parameters 0.6 B — https://huggingface.co/nvidia/parakeet-tdt-0.6b-v3 counted_at 2026-10-17T11:51:44.474Z — spec_sources.verified_at for total_params_b — when this figure was counted; a refresh that keeps it does not rewrite it confidence 0.95 — spec_sources.confidence — from the tensor headers The parameters of the language-model tensors only, counted from the tensor headers: a vision or audio tower and multi-token-prediction heads are excluded, so it can be below the whole-checkpoint count the formula's cross-check quotes. It sizes the weights only when no measured file size is stored for the chosen quantization; it is not what is read per token on a mixture-of-experts model. Shown to one decimal: the underlying tensor count is exact, but more digits would claim a precision the published figure does not have (SPEC §14).	Unavailable — No quantization of this model is in the catalog yet, so there is nothing to size its weights from. (missing: model_quantizations)	Unavailable — No quantization has been sized for this model, so the weights have no measured size to work from. (missing: no quantized repos listed (GGUF/MLX) — only the native checkpoint can be sized, Plan asks for quant "Q8_0" but no sized quantization of Parakeet TDT 0.6B v3 has that name, quantization file sizes)	Unavailable — No quantization has been sized for this model, so the weights have no measured size to work from. Nothing can be said about which machines run it until it can be sized. (missing: no quantized repos listed (GGUF/MLX) — only the native checkpoint can be sized, Plan asks for quant "Q8_0" but no sized quantization of Parakeet TDT 0.6B v3 has that name, quantization file sizes)	ONBOARDING — No quantization has been sized for this model, so the weights have no measured size to work from. Missing: no quantized repos listed (GGUF/MLX) — only the native checkpoint can be sized, Plan asks for quant "Q8_0" but no sized quantization of Parakeet TDT 0.6B v3 has that name, quantization file sizes.	read stale 3 d — past the 36 hour freshness window — Parakeet TDT 0.6B v3: the least recent read of its data from source (shape counted Oct 17, 7:51 AM ET from the tensor headers of https://huggingface.co/nvidia/parakeet-tdt-0.6b-v3; no quantization size stored); never checked upstream.
Whisper large-v3OpenAI · dense · speech to text	1.5 (derived) DERIVEDWhisper large-v3 total parametersArithmetic on observed inputs only. No tunable assumption enters it. Σ params of language-model tensors — Cross-checked: tensor count 1.54B vs Hugging Face 1.54B (0%). 1.5 B, counted from the tensor headers of https://huggingface.co/openai/whisper-large-v3 at 2026-10-17T11:54:13.096Z total_parameters 1.5 B — https://huggingface.co/openai/whisper-large-v3 counted_at 2026-10-17T11:54:13.096Z — spec_sources.verified_at for total_params_b — when this figure was counted; a refresh that keeps it does not rewrite it confidence 0.95 — spec_sources.confidence — from the tensor headers The parameters of the language-model tensors only, counted from the tensor headers: a vision or audio tower and multi-token-prediction heads are excluded, so it can be below the whole-checkpoint count the formula's cross-check quotes. It sizes the weights only when no measured file size is stored for the chosen quantization; it is not what is read per token on a mixture-of-experts model. Shown to one decimal: the underlying tensor count is exact, but more digits would claim a precision the published figure does not have (SPEC §14). 1.5 (derived) DERIVEDWhisper large-v3 active parameters per tokenArithmetic on observed inputs only. No tunable assumption enters it. active − input embedding table — Per-token memory read used for decode bandwidth (SPEC §8.2). 1.5 B, counted from the tensor headers of https://huggingface.co/openai/whisper-large-v3 at 2026-10-17T11:54:13.096Z active_parameters_per_token 1.5 B — https://huggingface.co/openai/whisper-large-v3 counted_at 2026-10-17T11:54:13.096Z — spec_sources.verified_at for active_params_b — when this figure was counted; a refresh that keeps it does not rewrite it confidence 0.95 — spec_sources.confidence — from the tensor headers The parameters read to produce one token. It sets the decode speed; on a dense model it equals the total. Shown to one decimal: the underlying tensor count is exact, but more digits would claim a precision the published figure does not have (SPEC §14). active	Unavailable — No quantization of this model is in the catalog yet, so there is nothing to size its weights from. (missing: model_quantizations)	Unavailable — No quantization has been sized for this model, so the weights have no measured size to work from. (missing: no quantized repos listed (GGUF/MLX) — only the native checkpoint can be sized, quantization file sizes)	Unavailable — No quantization has been sized for this model, so the weights have no measured size to work from. Nothing can be said about which machines run it until it can be sized. (missing: no quantized repos listed (GGUF/MLX) — only the native checkpoint can be sized, quantization file sizes)	ONBOARDING — No quantization has been sized for this model, so the weights have no measured size to work from. Missing: no quantized repos listed (GGUF/MLX) — only the native checkpoint can be sized, quantization file sizes.	read stale 3 d — past the 36 hour freshness window — Whisper large-v3: the least recent read of its data from source (shape counted Oct 17, 7:54 AM ET from the tensor headers of https://huggingface.co/openai/whisper-large-v3; no quantization size stored); never checked upstream.
Whisper large-v3-turboOpenAI · dense · speech to text	0.8 (derived) DERIVEDWhisper large-v3-turbo total parametersArithmetic on observed inputs only. No tunable assumption enters it. Σ params of language-model tensors — Cross-checked: tensor count 0.81B vs Hugging Face 0.81B (0%). 0.8 B, counted from the tensor headers of https://huggingface.co/openai/whisper-large-v3-turbo at 2026-10-17T11:54:54.001Z total_parameters 0.8 B — https://huggingface.co/openai/whisper-large-v3-turbo counted_at 2026-10-17T11:54:54.001Z — spec_sources.verified_at for total_params_b — when this figure was counted; a refresh that keeps it does not rewrite it confidence 0.95 — spec_sources.confidence — from the tensor headers The parameters of the language-model tensors only, counted from the tensor headers: a vision or audio tower and multi-token-prediction heads are excluded, so it can be below the whole-checkpoint count the formula's cross-check quotes. It sizes the weights only when no measured file size is stored for the chosen quantization; it is not what is read per token on a mixture-of-experts model. Shown to one decimal: the underlying tensor count is exact, but more digits would claim a precision the published figure does not have (SPEC §14). 0.7 (derived) DERIVEDWhisper large-v3-turbo active parameters per tokenArithmetic on observed inputs only. No tunable assumption enters it. active − input embedding table — Per-token memory read used for decode bandwidth (SPEC §8.2). 0.7 B, counted from the tensor headers of https://huggingface.co/openai/whisper-large-v3-turbo at 2026-10-17T11:54:54.001Z active_parameters_per_token 0.7 B — https://huggingface.co/openai/whisper-large-v3-turbo counted_at 2026-10-17T11:54:54.001Z — spec_sources.verified_at for active_params_b — when this figure was counted; a refresh that keeps it does not rewrite it confidence 0.95 — spec_sources.confidence — from the tensor headers The parameters read to produce one token. It sets the decode speed; on a dense model it equals the total. Shown to one decimal: the underlying tensor count is exact, but more digits would claim a precision the published figure does not have (SPEC §14). active	Unavailable — No quantization of this model is in the catalog yet, so there is nothing to size its weights from. (missing: model_quantizations)	Unavailable — No quantization has been sized for this model, so the weights have no measured size to work from. (missing: no quantized repos listed (GGUF/MLX) — only the native checkpoint can be sized, quantization file sizes)	Unavailable — No quantization has been sized for this model, so the weights have no measured size to work from. Nothing can be said about which machines run it until it can be sized. (missing: no quantized repos listed (GGUF/MLX) — only the native checkpoint can be sized, quantization file sizes)	ONBOARDING — No quantization has been sized for this model, so the weights have no measured size to work from. Missing: no quantized repos listed (GGUF/MLX) — only the native checkpoint can be sized, quantization file sizes.	read stale 3 d — past the 36 hour freshness window — Whisper large-v3-turbo: the least recent read of its data from source (shape counted Oct 17, 7:54 AM ET from the tensor headers of https://huggingface.co/openai/whisper-large-v3-turbo; no quantization size stored); never checked upstream.

The machine search reported: 1333 platform combinations were collapsed. For each socket and memory size the search carries the CHEAPEST build and, when it is a different part, the one with the most CPU compute (which is what prefill depends on); the rest differ only in price. 8 CPU+GPU hybrid candidates were the same parts as a GPU rig already in the list and are listed once. Memory = weights + KV cache at 32K + runtime overhead (SPEC §8.1); the weights come from the measured file size when Hugging Face has been read, and from bits-per-weight when it has not — each figure says which, inside it. “Machines” counts the catalog rigs that run the model at ≥ 15 tok/s; the denominator is every machine the search could build. A model with no shape contributes nothing to either column and says so rather than showing a zero.

What onboarding still owes

12 of 60 models

llm_models — 12 of 60 cannot be sized yet

Why it matters: A model without a shape or a measured quantization contributes no memory figure, so it cannot be placed on a machine, cannot enter a Plan's total, and cannot be compared with one that can. Every count on this screen excludes these 12 and says so.

How it gets filled: npm run worker -- onboard It reads each model's config and its repository file listing from Hugging Face and records the real sizes.

What onboarding still owes, grouped by the fields each group of models is missing
Still owed	Models	Which ones
no quantized repos listed (GGUF/MLX) — only the native checkpoint can be sized, quantization file sizes	9	name all 9 Canary-Qwen 2.5B, Chandra OCR 2, DeepSeek-OCR 2, Distil-Whisper distil-large-v3.5, dots.ocr, Granite Speech 4.1 2B, PaddleOCR-VL (0.9B), Whisper large-v3, Whisper large-v3-turbo
active_params_b, attention_params, bits per weight for bartowski/google_gemma-4-26B-A4B-it-GGUF:BF16 (model shape unknown), bits per weight for bartowski/google_gemma-4-26B-A4B-it-GGUF:IQ2_M (model shape unknown), bits per weight for bartowski/google_gemma-4-26B-A4B-it-GGUF:IQ2_S (model shape unknown), bits per weight for bartowski/google_gemma-4-26B-A4B-it-GGUF:IQ2_XS (model shape unknown), bits per weight for bartowski/google_gemma-4-26B-A4B-it-GGUF:IQ2_XXS (model shape unknown), bits per weight for bartowski/google_gemma-4-26B-A4B-it-GGUF:IQ3_M (model shape unknown), bits per weight for bartowski/google_gemma-4-26B-A4B-it-GGUF:IQ3_XS (model shape unknown), bits per weight for bartowski/google_gemma-4-26B-A4B-it-GGUF:IQ3_XXS (model shape unknown), bits per weight for bartowski/google_gemma-4-26B-A4B-it-GGUF:IQ4_NL (model shape unknown), bits per weight for bartowski/google_gemma-4-26B-A4B-it-GGUF:IQ4_XS (model shape unknown), bits per weight for bartowski/google_gemma-4-26B-A4B-it-GGUF:Q2_K (model shape unknown), bits per weight for bartowski/google_gemma-4-26B-A4B-it-GGUF:Q2_K_L (model shape unknown), bits per weight for bartowski/google_gemma-4-26B-A4B-it-GGUF:Q3_K_L (model shape unknown), bits per weight for bartowski/google_gemma-4-26B-A4B-it-GGUF:Q3_K_M (model shape unknown), bits per weight for bartowski/google_gemma-4-26B-A4B-it-GGUF:Q3_K_S (model shape unknown), bits per weight for bartowski/google_gemma-4-26B-A4B-it-GGUF:Q3_K_XL (model shape unknown), bits per weight for bartowski/google_gemma-4-26B-A4B-it-GGUF:Q4_0 (model shape unknown), bits per weight for bartowski/google_gemma-4-26B-A4B-it-GGUF:Q4_1 (model shape unknown), bits per weight for bartowski/google_gemma-4-26B-A4B-it-GGUF:Q4_K_L (model shape unknown), bits per weight for bartowski/google_gemma-4-26B-A4B-it-GGUF:Q4_K_M (model shape unknown), bits per weight for bartowski/google_gemma-4-26B-A4B-it-GGUF:Q4_K_S (model shape unknown), bits per weight for bartowski/google_gemma-4-26B-A4B-it-GGUF:Q5_K_L (model shape unknown), bits per weight for bartowski/google_gemma-4-26B-A4B-it-GGUF:Q5_K_M (model shape unknown), bits per weight for bartowski/google_gemma-4-26B-A4B-it-GGUF:Q5_K_S (model shape unknown), bits per weight for bartowski/google_gemma-4-26B-A4B-it-GGUF:Q6_K (model shape unknown), bits per weight for bartowski/google_gemma-4-26B-A4B-it-GGUF:Q6_K_L (model shape unknown), bits per weight for bartowski/google_gemma-4-26B-A4B-it-GGUF:Q8_0 (model shape unknown), bits per weight for ggml-org/gemma-4-26B-A4B-it-GGUF:BF16 (model shape unknown), bits per weight for ggml-org/gemma-4-26B-A4B-it-GGUF:Q4_0 (model shape unknown), bits per weight for ggml-org/gemma-4-26B-A4B-it-GGUF:Q8_0 (model shape unknown), bits per weight for lmstudio-community/gemma-4-26B-A4B-it-GGUF:Q4_K_M (model shape unknown), bits per weight for lmstudio-community/gemma-4-26B-A4B-it-GGUF:Q6_K (model shape unknown), bits per weight for lmstudio-community/gemma-4-26B-A4B-it-GGUF:Q8_0 (model shape unknown), bits per weight for lmstudio-community/gemma-4-26B-A4B-it-MLX-4bit:4bit (model shape unknown), bits per weight for lmstudio-community/gemma-4-26B-A4B-it-MLX-5bit:5bit (model shape unknown), bits per weight for lmstudio-community/gemma-4-26B-A4B-it-MLX-6bit:6bit (model shape unknown), bits per weight for lmstudio-community/gemma-4-26B-A4B-it-MLX-8bit:8bit (model shape unknown), bits per weight for mlx-community/gemma-4-26B-A4B-it-OptiQ-4bit:4bit (model shape unknown), bits per weight for mlx-community/gemma-4-26B-A4B-it-heretic-msq-2.6bit:2.6bit (model shape unknown), bits per weight for mlx-community/gemma-4-26b-a4b-it-4bit:4bit (model shape unknown), bits per weight for mlx-community/gemma-4-26b-a4b-it-6bit:6bit (model shape unknown), bits per weight for mlx-community/gemma-4-26b-a4b-it-8bit:8bit (model shape unknown), bits per weight for mlx-community/gemma-4-26b-a4b-it-mxfp8:mxfp8 (model shape unknown), bits per weight for mlx-community/gemma-4-26b-a4b-it-nvfp4:native (model shape unknown), bits per weight for unsloth/gemma-4-26B-A4B-it-GGUF:BF16 (model shape unknown), bits per weight for unsloth/gemma-4-26B-A4B-it-GGUF:F16 (model shape unknown), bits per weight for unsloth/gemma-4-26B-A4B-it-GGUF:MXFP4_MOE (model shape unknown), bits per weight for unsloth/gemma-4-26B-A4B-it-GGUF:Q8_0 (model shape unknown), bits per weight for unsloth/gemma-4-26B-A4B-it-GGUF:UD-IQ2_M (model shape unknown), bits per weight for unsloth/gemma-4-26B-A4B-it-GGUF:UD-IQ2_XXS (model shape unknown), bits per weight for unsloth/gemma-4-26B-A4B-it-GGUF:UD-IQ3_S (model shape unknown), bits per weight for unsloth/gemma-4-26B-A4B-it-GGUF:UD-IQ3_XXS (model shape unknown), bits per weight for unsloth/gemma-4-26B-A4B-it-GGUF:UD-IQ4_NL (model shape unknown), bits per weight for unsloth/gemma-4-26B-A4B-it-GGUF:UD-IQ4_XS (model shape unknown), bits per weight for unsloth/gemma-4-26B-A4B-it-GGUF:UD-Q2_K_XL (model shape unknown), bits per weight for unsloth/gemma-4-26B-A4B-it-GGUF:UD-Q3_K_M (model shape unknown), bits per weight for unsloth/gemma-4-26B-A4B-it-GGUF:UD-Q3_K_XL (model shape unknown), bits per weight for unsloth/gemma-4-26B-A4B-it-GGUF:UD-Q4_K_M (model shape unknown), bits per weight for unsloth/gemma-4-26B-A4B-it-GGUF:UD-Q4_K_S (model shape unknown), bits per weight for unsloth/gemma-4-26B-A4B-it-GGUF:UD-Q4_K_XL (model shape unknown), bits per weight for unsloth/gemma-4-26B-A4B-it-GGUF:UD-Q5_K_M (model shape unknown), bits per weight for unsloth/gemma-4-26B-A4B-it-GGUF:UD-Q5_K_S (model shape unknown), bits per weight for unsloth/gemma-4-26B-A4B-it-GGUF:UD-Q5_K_XL (model shape unknown), bits per weight for unsloth/gemma-4-26B-A4B-it-GGUF:UD-Q6_K (model shape unknown), bits per weight for unsloth/gemma-4-26B-A4B-it-GGUF:UD-Q6_K_XL (model shape unknown), bits per weight for unsloth/gemma-4-26B-A4B-it-GGUF:UD-Q8_K_XL (model shape unknown), bits per weight for unsloth/gemma-4-26B-A4B-it-NVFP4:native (model shape unknown), bits per weight for unsloth/gemma-4-26b-a4b-it-MLX-8bit:8bit (model shape unknown), bits per weight for unsloth/gemma-4-26b-a4b-it-UD-MLX-3bit:3bit (model shape unknown), bits per weight for unsloth/gemma-4-26b-a4b-it-UD-MLX-4bit:4bit (model shape unknown), model shape (tensor inventory and config formula both failed), num_layers, total_params_b	1	name the model Gemma 4 26B-A4B IT
Plan asks for quant "4bit" but no sized quantization of DeepSeek-V4.1-Flash has that name, no quantized repos listed (GGUF/MLX) — only the native checkpoint can be sized, quantization file sizes	1	name the model DeepSeek-V4.1-Flash
Plan asks for quant "Q8_0" but no sized quantization of Parakeet TDT 0.6B v3 has that name, no quantized repos listed (GGUF/MLX) — only the native checkpoint can be sized, quantization file sizes	1	name the model Parakeet TDT 0.6B v3

Qwen3-VL 235B-A22B Instruct

Alibaba Qwen · vision · needs 150 GB at 32K ctx · runs on 14 of 89 machines at ≥ 15 tok/s · 53 quantizations · 2 published results

What it needs, on any machine

Weights

133 GB

OBSERVEDQwen3-VL 235B-A22B Instruct weights at IQ4_NLMeasured or published, passed through unchanged — not computed here.

weights_gb = real file size (Hugging Face API)

weights_gb = 132.6 GB (IQ4_NL)

file_size_gb: 132.60341664 GB — https://huggingface.co/unsloth/Qwen3-VL-235B-A22B-Instruct-1M-GGUF/tree/main

Real file size used instead of the params × bits formula (SPEC §8.1).

KV cache

6.31 GB

DERIVEDQwen3-VL 235B-A22B Instruct KV cache at 32K contextArithmetic on observed inputs only. No tunable assumption enters it.

kv_cache_gb = Σ over full-attention layers of per-token cache width × kv_bytes (from tensor shapes) × context / 1e9

kv_cache_gb = 96,256 elements × 2 × 32,768 / 1e9 = 6.31 GB

num_layers: 94 — model config
context: 32768 tokens — Plan context setting
kv_bytes: 2 bytes — Settings (fp16 cache)

94 layers keep a full KV cache.

Overhead

10.6 GB

ESTIMATEQwen3-VL 235B-A22B Instruct runtime overheadArithmetic that also depends on a tunable assumption, listed below.

overhead_gb = weights_gb × overhead_fraction

overhead_gb = 132.6 × 0.08 = 10.61 GB

overhead_fraction: 0.08 — Settings (SPEC §8.1 default 0.08)

Runtime buffers, activations, and allocator slack approximated as a fraction of weights.

Total

150 GB

ESTIMATEQwen3-VL 235B-A22B Instruct total memory requirementArithmetic that also depends on a tunable assumption, listed below.

required_gb = weights_gb + kv_cache_gb + overhead_gb

required_gb = 132.6 + 6.31 + 10.61 = 149.52 GB

weights_gb: 132.60341664 GB — real file size
kv_cache_gb: 6.308233216 GB — kv formula
overhead_gb: 10.608273331200001 GB — overhead fraction

Real file size used instead of the params × bits formula (SPEC §8.1).
94 layers keep a full KV cache.
Runtime buffers, activations, and allocator slack approximated as a fraction of weights.

The weights figure is the measured IQ4_NL download, not a bits-per-weight estimate. This requirement is a property of the model and the context — it is the same on every machine.

Where this model’s facts come from

Repository

Qwen/Qwen3-VL-235B-A22B-Instruct

Licence

apache-2.0 (permissive)

Released

2025-09-22

Maximum context

262,144

OBSERVEDQwen3-VL 235B-A22B Instruct published maximum contextMeasured or published, passed through unchanged — not computed here.

published maximum context — published at the source, passed through unchanged; nothing is computed here

262,144 tokens, as recorded from max_position_embeddings in the config.json of Qwen/Qwen3-VL-235B-A22B-Instruct

published_maximum_context: 262,144 tokens — max_position_embeddings in the config.json of Qwen/Qwen3-VL-235B-A22B-Instruct

What the maintainers support. A longer context can often be forced; this app sizes the KV cache for the context you ask for and does not clamp it.
Read by the models stage's refresh. When that read happened is not recorded apart from seed loads (llm_models.verified_at moves on both), so no read date is claimed for it.
Observed, not derived: this app records the figure and its source, and does not re-measure it.

tokens

Shape counted

Oct 17, 8:11 AM ET (3 d ago) from the tensor headers of https://huggingface.co/Qwen/Qwen3-VL-235B-A22B-Instruct — kept, not re-counted, while the repository has no new commit, and re-counted at least every 14 days

Quantization sizes read

Oct 17, 8:11 AM ET (3 d ago) — the listings behind all 53 of its quantization rows, re-read with every refresh

Checked upstream

Current as of

stale 3 d — the least recent read from source, the newer of the two; stale past the 36 h freshness window in Settings

Row last written

Oct 17, 10:27 AM ET (3 d ago) — llm_models.verified_at, moved by every seed load and every refresh — not a read from source, and it dates none of the figures above

Onboarding

complete · via seed

Quantizations, with the size measured on Hugging Face

Size this model with

Showing the default pick: the measured quantization closest to 4.5 bits per weight.

10 of the 53 quantizations of Qwen3-VL 235B-A22B Instruct in the catalog, with the file size measured on Hugging Face
Quantization	Bits/weight	File size GB	Quality Δ	Source
IQ4_XSgguf · unsloth/Qwen3-VL-235B-A22B-Instruct-1M-GGUF	4.26 (observed) OBSERVEDIQ4_XS bits per weightMeasured or published, passed through unchanged — not computed here. bits per weight — published at the source, passed through unchanged; nothing is computed here 4.26 bits/weight, as recorded from https://huggingface.co/unsloth/Qwen3-VL-235B-A22B-Instruct-1M-GGUF/tree/main bits_per_weight 4.26 bits/weight — https://huggingface.co/unsloth/Qwen3-VL-235B-A22B-Instruct-1M-GGUF/tree/main verified_at 2026-10-17T12:11:06.530Z — the row's own verification timestamp An average over the checkpoint: mixed quantizations keep some tensors at higher precision, which this figure already reflects. Observed, not derived: this app records the figure and its source, and does not re-measure it.	125.30 (observed) OBSERVEDIQ4_XS file size on diskMeasured or published, passed through unchanged — not computed here. quantization file size — published at the source, passed through unchanged; nothing is computed here 125.30 GB, as recorded from https://huggingface.co/unsloth/Qwen3-VL-235B-A22B-Instruct-1M-GGUF/tree/main, file 3 files quantization_file_size 125.30 GB — https://huggingface.co/unsloth/Qwen3-VL-235B-A22B-Instruct-1M-GGUF/tree/main, file 3 files verified_at 2026-10-17T12:11:06.530Z — the row's own verification timestamp quantization IQ4_XS — gguf bits_per_weight 4.2639 bits — the quantizer's own naming, or measured from the file The on-disk size of the weights. It is not the memory the model occupies: add the KV cache and the runtime overhead (SPEC §8.1). Preferred over the bits-per-weight formula wherever it exists, because it includes embeddings and any higher-precision tensors. Observed, not derived: this app records the figure and its source, and does not re-measure it.	Unavailable — No published quality measurement for IQ4_XS, so the accuracy cost of choosing this quantization over the full-precision checkpoint is not known. (missing: quality_delta_pct)	3 filesverified Oct 17, 8:11 AM ET
IQ4_XSgguf · unsloth/Qwen3-VL-235B-A22B-Instruct-GGUF	4.26 (observed) OBSERVEDIQ4_XS bits per weightMeasured or published, passed through unchanged — not computed here. bits per weight — published at the source, passed through unchanged; nothing is computed here 4.26 bits/weight, as recorded from https://huggingface.co/unsloth/Qwen3-VL-235B-A22B-Instruct-GGUF/tree/main bits_per_weight 4.26 bits/weight — https://huggingface.co/unsloth/Qwen3-VL-235B-A22B-Instruct-GGUF/tree/main verified_at 2026-10-17T12:11:06.530Z — the row's own verification timestamp An average over the checkpoint: mixed quantizations keep some tensors at higher precision, which this figure already reflects. Observed, not derived: this app records the figure and its source, and does not re-measure it.	125.30 (observed) OBSERVEDIQ4_XS file size on diskMeasured or published, passed through unchanged — not computed here. quantization file size — published at the source, passed through unchanged; nothing is computed here 125.30 GB, as recorded from https://huggingface.co/unsloth/Qwen3-VL-235B-A22B-Instruct-GGUF/tree/main, file 3 files quantization_file_size 125.30 GB — https://huggingface.co/unsloth/Qwen3-VL-235B-A22B-Instruct-GGUF/tree/main, file 3 files verified_at 2026-10-17T12:11:06.530Z — the row's own verification timestamp quantization IQ4_XS — gguf bits_per_weight 4.2639 bits — the quantizer's own naming, or measured from the file The on-disk size of the weights. It is not the memory the model occupies: add the KV cache and the runtime overhead (SPEC §8.1). Preferred over the bits-per-weight formula wherever it exists, because it includes embeddings and any higher-precision tensors. Observed, not derived: this app records the figure and its source, and does not re-measure it.	Unavailable — No published quality measurement for IQ4_XS, so the accuracy cost of choosing this quantization over the full-precision checkpoint is not known. (missing: quality_delta_pct)	3 filesverified Oct 17, 8:11 AM ET
Currently selected: IQ4_NLgguf · unsloth/Qwen3-VL-235B-A22B-Instruct-1M-GGUF	4.51 (observed) OBSERVEDIQ4_NL bits per weightMeasured or published, passed through unchanged — not computed here. bits per weight — published at the source, passed through unchanged; nothing is computed here 4.51 bits/weight, as recorded from https://huggingface.co/unsloth/Qwen3-VL-235B-A22B-Instruct-1M-GGUF/tree/main bits_per_weight 4.51 bits/weight — https://huggingface.co/unsloth/Qwen3-VL-235B-A22B-Instruct-1M-GGUF/tree/main verified_at 2026-10-17T12:11:06.530Z — the row's own verification timestamp An average over the checkpoint: mixed quantizations keep some tensors at higher precision, which this figure already reflects. Observed, not derived: this app records the figure and its source, and does not re-measure it.	132.60 (observed) OBSERVEDIQ4_NL file size on diskMeasured or published, passed through unchanged — not computed here. quantization file size — published at the source, passed through unchanged; nothing is computed here 132.60 GB, as recorded from https://huggingface.co/unsloth/Qwen3-VL-235B-A22B-Instruct-1M-GGUF/tree/main, file 3 files quantization_file_size 132.60 GB — https://huggingface.co/unsloth/Qwen3-VL-235B-A22B-Instruct-1M-GGUF/tree/main, file 3 files verified_at 2026-10-17T12:11:06.530Z — the row's own verification timestamp quantization IQ4_NL — gguf bits_per_weight 4.5124 bits — the quantizer's own naming, or measured from the file The on-disk size of the weights. It is not the memory the model occupies: add the KV cache and the runtime overhead (SPEC §8.1). Preferred over the bits-per-weight formula wherever it exists, because it includes embeddings and any higher-precision tensors. Observed, not derived: this app records the figure and its source, and does not re-measure it.	Unavailable — No published quality measurement for IQ4_NL, so the accuracy cost of choosing this quantization over the full-precision checkpoint is not known. (missing: quality_delta_pct)	3 filesverified Oct 17, 8:11 AM ET
IQ4_NLgguf · unsloth/Qwen3-VL-235B-A22B-Instruct-GGUF	4.51 (observed) OBSERVEDIQ4_NL bits per weightMeasured or published, passed through unchanged — not computed here. bits per weight — published at the source, passed through unchanged; nothing is computed here 4.51 bits/weight, as recorded from https://huggingface.co/unsloth/Qwen3-VL-235B-A22B-Instruct-GGUF/tree/main bits_per_weight 4.51 bits/weight — https://huggingface.co/unsloth/Qwen3-VL-235B-A22B-Instruct-GGUF/tree/main verified_at 2026-10-17T12:11:06.530Z — the row's own verification timestamp An average over the checkpoint: mixed quantizations keep some tensors at higher precision, which this figure already reflects. Observed, not derived: this app records the figure and its source, and does not re-measure it.	132.60 (observed) OBSERVEDIQ4_NL file size on diskMeasured or published, passed through unchanged — not computed here. quantization file size — published at the source, passed through unchanged; nothing is computed here 132.60 GB, as recorded from https://huggingface.co/unsloth/Qwen3-VL-235B-A22B-Instruct-GGUF/tree/main, file 3 files quantization_file_size 132.60 GB — https://huggingface.co/unsloth/Qwen3-VL-235B-A22B-Instruct-GGUF/tree/main, file 3 files verified_at 2026-10-17T12:11:06.530Z — the row's own verification timestamp quantization IQ4_NL — gguf bits_per_weight 4.5124 bits — the quantizer's own naming, or measured from the file The on-disk size of the weights. It is not the memory the model occupies: add the KV cache and the runtime overhead (SPEC §8.1). Preferred over the bits-per-weight formula wherever it exists, because it includes embeddings and any higher-precision tensors. Observed, not derived: this app records the figure and its source, and does not re-measure it.	Unavailable — No published quality measurement for IQ4_NL, so the accuracy cost of choosing this quantization over the full-precision checkpoint is not known. (missing: quality_delta_pct)	3 filesverified Oct 17, 8:11 AM ET
Q4_0gguf · unsloth/Qwen3-VL-235B-A22B-Instruct-1M-GGUF	4.53 (observed) OBSERVEDQ4_0 bits per weightMeasured or published, passed through unchanged — not computed here. bits per weight — published at the source, passed through unchanged; nothing is computed here 4.53 bits/weight, as recorded from https://huggingface.co/unsloth/Qwen3-VL-235B-A22B-Instruct-1M-GGUF/tree/main bits_per_weight 4.53 bits/weight — https://huggingface.co/unsloth/Qwen3-VL-235B-A22B-Instruct-1M-GGUF/tree/main verified_at 2026-10-17T12:11:06.530Z — the row's own verification timestamp An average over the checkpoint: mixed quantizations keep some tensors at higher precision, which this figure already reflects. Observed, not derived: this app records the figure and its source, and does not re-measure it.	133.13 (observed) OBSERVEDQ4_0 file size on diskMeasured or published, passed through unchanged — not computed here. quantization file size — published at the source, passed through unchanged; nothing is computed here 133.13 GB, as recorded from https://huggingface.co/unsloth/Qwen3-VL-235B-A22B-Instruct-1M-GGUF/tree/main, file 3 files quantization_file_size 133.13 GB — https://huggingface.co/unsloth/Qwen3-VL-235B-A22B-Instruct-1M-GGUF/tree/main, file 3 files verified_at 2026-10-17T12:11:06.530Z — the row's own verification timestamp quantization Q4_0 — gguf bits_per_weight 4.5304 bits — the quantizer's own naming, or measured from the file The on-disk size of the weights. It is not the memory the model occupies: add the KV cache and the runtime overhead (SPEC §8.1). Preferred over the bits-per-weight formula wherever it exists, because it includes embeddings and any higher-precision tensors. Observed, not derived: this app records the figure and its source, and does not re-measure it.	Unavailable — No published quality measurement for Q4_0, so the accuracy cost of choosing this quantization over the full-precision checkpoint is not known. (missing: quality_delta_pct)	3 filesverified Oct 17, 8:11 AM ET
Q4_0gguf · unsloth/Qwen3-VL-235B-A22B-Instruct-GGUF	4.53 (observed) OBSERVEDQ4_0 bits per weightMeasured or published, passed through unchanged — not computed here. bits per weight — published at the source, passed through unchanged; nothing is computed here 4.53 bits/weight, as recorded from https://huggingface.co/unsloth/Qwen3-VL-235B-A22B-Instruct-GGUF/tree/main bits_per_weight 4.53 bits/weight — https://huggingface.co/unsloth/Qwen3-VL-235B-A22B-Instruct-GGUF/tree/main verified_at 2026-10-17T12:11:06.530Z — the row's own verification timestamp An average over the checkpoint: mixed quantizations keep some tensors at higher precision, which this figure already reflects. Observed, not derived: this app records the figure and its source, and does not re-measure it.	133.13 (observed) OBSERVEDQ4_0 file size on diskMeasured or published, passed through unchanged — not computed here. quantization file size — published at the source, passed through unchanged; nothing is computed here 133.13 GB, as recorded from https://huggingface.co/unsloth/Qwen3-VL-235B-A22B-Instruct-GGUF/tree/main, file 3 files quantization_file_size 133.13 GB — https://huggingface.co/unsloth/Qwen3-VL-235B-A22B-Instruct-GGUF/tree/main, file 3 files verified_at 2026-10-17T12:11:06.530Z — the row's own verification timestamp quantization Q4_0 — gguf bits_per_weight 4.5304 bits — the quantizer's own naming, or measured from the file The on-disk size of the weights. It is not the memory the model occupies: add the KV cache and the runtime overhead (SPEC §8.1). Preferred over the bits-per-weight formula wherever it exists, because it includes embeddings and any higher-precision tensors. Observed, not derived: this app records the figure and its source, and does not re-measure it.	Unavailable — No published quality measurement for Q4_0, so the accuracy cost of choosing this quantization over the full-precision checkpoint is not known. (missing: quality_delta_pct)	3 filesverified Oct 17, 8:11 AM ET
Q4_K_Sgguf · unsloth/Qwen3-VL-235B-A22B-Instruct-1M-GGUF	4.55 (observed) OBSERVEDQ4_K_S bits per weightMeasured or published, passed through unchanged — not computed here. bits per weight — published at the source, passed through unchanged; nothing is computed here 4.55 bits/weight, as recorded from https://huggingface.co/unsloth/Qwen3-VL-235B-A22B-Instruct-1M-GGUF/tree/main bits_per_weight 4.55 bits/weight — https://huggingface.co/unsloth/Qwen3-VL-235B-A22B-Instruct-1M-GGUF/tree/main verified_at 2026-10-17T12:11:06.530Z — the row's own verification timestamp An average over the checkpoint: mixed quantizations keep some tensors at higher precision, which this figure already reflects. Observed, not derived: this app records the figure and its source, and does not re-measure it.	133.69 (observed) OBSERVEDQ4_K_S file size on diskMeasured or published, passed through unchanged — not computed here. quantization file size — published at the source, passed through unchanged; nothing is computed here 133.69 GB, as recorded from https://huggingface.co/unsloth/Qwen3-VL-235B-A22B-Instruct-1M-GGUF/tree/main, file 3 files quantization_file_size 133.69 GB — https://huggingface.co/unsloth/Qwen3-VL-235B-A22B-Instruct-1M-GGUF/tree/main, file 3 files verified_at 2026-10-17T12:11:06.530Z — the row's own verification timestamp quantization Q4_K_S — gguf bits_per_weight 4.5492 bits — the quantizer's own naming, or measured from the file The on-disk size of the weights. It is not the memory the model occupies: add the KV cache and the runtime overhead (SPEC §8.1). Preferred over the bits-per-weight formula wherever it exists, because it includes embeddings and any higher-precision tensors. Observed, not derived: this app records the figure and its source, and does not re-measure it.	Unavailable — No published quality measurement for Q4_K_S, so the accuracy cost of choosing this quantization over the full-precision checkpoint is not known. (missing: quality_delta_pct)	3 filesverified Oct 17, 8:11 AM ET
Q4_K_Sgguf · unsloth/Qwen3-VL-235B-A22B-Instruct-GGUF	4.55 (observed) OBSERVEDQ4_K_S bits per weightMeasured or published, passed through unchanged — not computed here. bits per weight — published at the source, passed through unchanged; nothing is computed here 4.55 bits/weight, as recorded from https://huggingface.co/unsloth/Qwen3-VL-235B-A22B-Instruct-GGUF/tree/main bits_per_weight 4.55 bits/weight — https://huggingface.co/unsloth/Qwen3-VL-235B-A22B-Instruct-GGUF/tree/main verified_at 2026-10-17T12:11:06.530Z — the row's own verification timestamp An average over the checkpoint: mixed quantizations keep some tensors at higher precision, which this figure already reflects. Observed, not derived: this app records the figure and its source, and does not re-measure it.	133.69 (observed) OBSERVEDQ4_K_S file size on diskMeasured or published, passed through unchanged — not computed here. quantization file size — published at the source, passed through unchanged; nothing is computed here 133.69 GB, as recorded from https://huggingface.co/unsloth/Qwen3-VL-235B-A22B-Instruct-GGUF/tree/main, file 3 files quantization_file_size 133.69 GB — https://huggingface.co/unsloth/Qwen3-VL-235B-A22B-Instruct-GGUF/tree/main, file 3 files verified_at 2026-10-17T12:11:06.530Z — the row's own verification timestamp quantization Q4_K_S — gguf bits_per_weight 4.5492 bits — the quantizer's own naming, or measured from the file The on-disk size of the weights. It is not the memory the model occupies: add the KV cache and the runtime overhead (SPEC §8.1). Preferred over the bits-per-weight formula wherever it exists, because it includes embeddings and any higher-precision tensors. Observed, not derived: this app records the figure and its source, and does not re-measure it.	Unavailable — No published quality measurement for Q4_K_S, so the accuracy cost of choosing this quantization over the full-precision checkpoint is not known. (missing: quality_delta_pct)	3 filesverified Oct 17, 8:11 AM ET
UD-Q4_K_XLgguf · unsloth/Qwen3-VL-235B-A22B-Instruct-1M-GGUF	4.56 (observed) OBSERVEDUD-Q4_K_XL bits per weightMeasured or published, passed through unchanged — not computed here. bits per weight — published at the source, passed through unchanged; nothing is computed here 4.56 bits/weight, as recorded from https://huggingface.co/unsloth/Qwen3-VL-235B-A22B-Instruct-1M-GGUF/tree/main bits_per_weight 4.56 bits/weight — https://huggingface.co/unsloth/Qwen3-VL-235B-A22B-Instruct-1M-GGUF/tree/main verified_at 2026-10-17T12:11:06.530Z — the row's own verification timestamp An average over the checkpoint: mixed quantizations keep some tensors at higher precision, which this figure already reflects. Observed, not derived: this app records the figure and its source, and does not re-measure it.	134.12 (observed) OBSERVEDUD-Q4_K_XL file size on diskMeasured or published, passed through unchanged — not computed here. quantization file size — published at the source, passed through unchanged; nothing is computed here 134.12 GB, as recorded from https://huggingface.co/unsloth/Qwen3-VL-235B-A22B-Instruct-1M-GGUF/tree/main, file 3 files quantization_file_size 134.12 GB — https://huggingface.co/unsloth/Qwen3-VL-235B-A22B-Instruct-1M-GGUF/tree/main, file 3 files verified_at 2026-10-17T12:11:06.530Z — the row's own verification timestamp quantization UD-Q4_K_XL — gguf bits_per_weight 4.5641 bits — the quantizer's own naming, or measured from the file The on-disk size of the weights. It is not the memory the model occupies: add the KV cache and the runtime overhead (SPEC §8.1). Preferred over the bits-per-weight formula wherever it exists, because it includes embeddings and any higher-precision tensors. Observed, not derived: this app records the figure and its source, and does not re-measure it.	Unavailable — No published quality measurement for UD-Q4_K_XL, so the accuracy cost of choosing this quantization over the full-precision checkpoint is not known. (missing: quality_delta_pct)	3 filesverified Oct 17, 8:11 AM ET
UD-Q4_K_XLgguf · unsloth/Qwen3-VL-235B-A22B-Instruct-GGUF	4.56 (observed) OBSERVEDUD-Q4_K_XL bits per weightMeasured or published, passed through unchanged — not computed here. bits per weight — published at the source, passed through unchanged; nothing is computed here 4.56 bits/weight, as recorded from https://huggingface.co/unsloth/Qwen3-VL-235B-A22B-Instruct-GGUF/tree/main bits_per_weight 4.56 bits/weight — https://huggingface.co/unsloth/Qwen3-VL-235B-A22B-Instruct-GGUF/tree/main verified_at 2026-10-17T12:11:06.530Z — the row's own verification timestamp An average over the checkpoint: mixed quantizations keep some tensors at higher precision, which this figure already reflects. Observed, not derived: this app records the figure and its source, and does not re-measure it.	134.12 (observed) OBSERVEDUD-Q4_K_XL file size on diskMeasured or published, passed through unchanged — not computed here. quantization file size — published at the source, passed through unchanged; nothing is computed here 134.12 GB, as recorded from https://huggingface.co/unsloth/Qwen3-VL-235B-A22B-Instruct-GGUF/tree/main, file 3 files quantization_file_size 134.12 GB — https://huggingface.co/unsloth/Qwen3-VL-235B-A22B-Instruct-GGUF/tree/main, file 3 files verified_at 2026-10-17T12:11:06.530Z — the row's own verification timestamp quantization UD-Q4_K_XL — gguf bits_per_weight 4.5641 bits — the quantizer's own naming, or measured from the file The on-disk size of the weights. It is not the memory the model occupies: add the KV cache and the runtime overhead (SPEC §8.1). Preferred over the bits-per-weight formula wherever it exists, because it includes embeddings and any higher-precision tensors. Observed, not derived: this app records the figure and its source, and does not re-measure it.	Unavailable — No published quality measurement for UD-Q4_K_XL, so the accuracy cost of choosing this quantization over the full-precision checkpoint is not known. (missing: quality_delta_pct)	3 filesverified Oct 17, 8:11 AM ET

Showing the 10 quantizations nearest 4.51 bits per weight — the realistic alternatives to the one selected. 43 more are in the catalog, from 3 repositories and 1 formats — show all 53.

Machines in the catalog that run it

Listed: every machine that can hold Qwen3-VL 235B-A22B Instruct in memory. 14 of these 54 also clear ≥ 15 tok/s, the speed floor the saved Plan asks for. 35 more machines in the catalog cannot hold this model at all and are not listed. Prices are the listed price of the machine, before shipping and sales tax — the landed cost, and the shipping that is missing from it, are on Compare.

Published results

Published benchmark results on record for Qwen3-VL 235B-A22B Instruct
Benchmark	Score	What was measured, and how
MMMU-Pro (Artificial Analysis run)	67.60 (observed) OBSERVEDMMMU-Pro (Artificial Analysis run) scoreMeasured or published, passed through unchanged — not computed here. MMMU-Pro (Artificial Analysis run) — published at the source, passed through unchanged; nothing is computed here 67.6, as recorded from https://artificialanalysis.ai/models/qwen3-vl-235b-a22b-instruct MMMU-Pro_(Artificial_Analysis_run) 67.6 — https://artificialanalysis.ai/models/qwen3-vl-235b-a22b-instruct measured % of MMMU-Pro image questions answered correctly (30 academic disciplines); Artificial Analysis's own run of "Qwen3 VL 235B A22B Instruct" through an API endpoint (not local quantized inference); AA payload value 0.6757; AA lists this model as deprecated (successor: qwen3-5-397b-a17b) but keeps its scores; read 2026-10-17 from the AA model page Measured on this model, as the source states. Benchmarks are run at full precision unless the metric says otherwise; a quantized local copy will not match them exactly. Recorded in the seed file (seed/models.json) with its source URL but no date of reading, so none is claimed; the model row's write time is a seed load, not a reading of this source. Observed, not derived: this app records the figure and its source, and does not re-measure it.	% of MMMU-Pro image questions answered correctly (30 academic disciplines); Artificial Analysis's own run of "Qwen3 VL 235B A22B Instruct" through an API endpoint (not local quantized inference); AA payload value 0.6757; AA lists this model as deprecated (successor: qwen3-5-397b-a17b) but keeps its scores; read 2026-10-17 from the AA model pagesource
OmniDocBench v1.6 end-to-end Overall	89.78 (observed) OBSERVEDOmniDocBench v1.6 end-to-end Overall scoreMeasured or published, passed through unchanged — not computed here. OmniDocBench v1.6 end-to-end Overall — published at the source, passed through unchanged; nothing is computed here 89.78, as recorded from https://github.com/opendatalab/OmniDocBench/blob/f133a71e9e91c3621c7ce8994200a7b394a06eb3/README.md OmniDocBench_v1.6_end-to-end_Overall 89.78 — https://github.com/opendatalab/OmniDocBench/blob/f133a71e9e91c3621c7ce8994200a7b394a06eb3/README.md measured Overall document-parsing score 0-100, higher is better: ((1 - text edit distance) x 100 + table TEDS + formula CDM) / 3; row "Qwen3-VL-235B" of the README table 'Comprehensive evaluation of document parsing on OmniDocBench (v1.6_full)', evaluated weights Qwen/Qwen3-VL-235B-A22B-Instruct per the README's model-information table; leaderboard published by the OmniDocBench maintainers (OpenDataLab, Shanghai AI Lab), README commit f133a71e9e of 2026-09-11; read 2026-10-17 Measured on this model, as the source states. Benchmarks are run at full precision unless the metric says otherwise; a quantized local copy will not match them exactly. Recorded in the seed file (seed/models.json) with its source URL but no date of reading, so none is claimed; the model row's write time is a seed load, not a reading of this source. Observed, not derived: this app records the figure and its source, and does not re-measure it.	Overall document-parsing score 0-100, higher is better: ((1 - text edit distance) x 100 + table TEDS + formula CDM) / 3; row "Qwen3-VL-235B" of the README table 'Comprehensive evaluation of document parsing on OmniDocBench (v1.6_full)', evaluated weights Qwen/Qwen3-VL-235B-A22B-Instruct per the README's model-information table; leaderboard published by the OmniDocBench maintainers (OpenDataLab, Shanghai AI Lab), README commit f133a71e9e of 2026-09-11; read 2026-10-17source

The 54 machines in the catalog that can hold Qwen3-VL 235B-A22B Instruct in memory, fastest first
Machine	Verdict	Decode	TTFT 32K	Observed price
NVIDIA DGX Station GB300 (Grace Blackwell Ultra), 748GB 691 GB (derived) DERIVEDNVIDIA DGX Station GB300 (Grace Blackwell Ultra), 748GB usable memoryArithmetic on observed inputs only. No tunable assumption enters it. available_gb = Σ usable_gb over memory pools GPU HBM3e 244.44 + CPU LPDDR5X 446.4 = 690.84 GB GPU HBM3e.usable_gb 244.44 GB — 252 GB physical CPU LPDDR5X.usable_gb 446.40000000000003 GB — 496 GB physical Usable memory is below physical memory for at least one pool (OS/driver reservation; editable in Settings). usable 7,100 (observed) OBSERVEDNVIDIA DGX Station GB300 (Grace Blackwell Ultra), 748GB peak memory bandwidthMeasured or published, passed through unchanged — not computed here. peak memory bandwidth — published at the source, passed through unchanged; nothing is computed here 7,100 GB/s, as recorded from the GPU HBM3e spec of NVIDIA DGX Station GB300 (Grace Blackwell Ultra), 748GB peak_memory_bandwidth 7,100 GB/s — the GPU HBM3e spec of NVIDIA DGX Station GB300 (Grace Blackwell Ultra), 748GB Theoretical peak. Decode reaches a fraction of it — the efficiency factor is in the decode speed's own workings. The fastest pool. A machine with more than one pool places weights fastest-pool-first. Observed, not derived: this app records the figure and its source, and does not re-measure it. GB/s	Fast	143 tok/s (estimate) ESTIMATEQwen3-VL 235B-A22B Instruct decode speed on NVIDIA DGX Station GB300 (Grace Blackwell Ultra), 748GBArithmetic that also depends on a tunable assumption, listed below. tok_s = 1 / (bytes_per_token / (bandwidth × efficiency) + overhead) GPU HBM3e: 18.474 GB ÷ (7,100 GB/s × 0.48) + 1.58 ms fixed per-token overhead → 142.84 tok/s GPU HBM3e.bandwidth_gbs 7100 GB/s — component spec GPU HBM3e.efficiency_moe 0.48 — Settings efficiency.gpu_vram — Discrete GPU: 10 dense + 3 MoE points; fit on published benchmarks, leave-one-out validated — docs/research/12b-estimator-validation.md GPU HBM3e.active_share 1 — placement per_token_overhead_ms 1.58 ms — Settings efficiency.gpu_vram.overheadMs — calibrated fit (D-013): Discrete GPU: 10 dense + 3 MoE points; fit on published benchmarks, leave-one-out validated — docs/research/12b-estimator-validation.md bytes_per_token_gb 18.4738 GB MoE: only active parameters are read per token (SPEC §3). Effective bits include embeddings and higher-precision tensors in mixed quants. Assumes active weights are quantized like the file average. KV cache is re-read on every token (DECISIONS D-008). Batch size 1 (single user, interactive).	808 ms (estimate) ESTIMATEQwen3-VL 235B-A22B Instruct time to first token on NVIDIA DGX Station GB300 (Grace Blackwell Ultra), 748GBArithmetic that also depends on a tunable assumption, listed below. ttft_s = prompt_tokens / prefill_tok_s = 32,768 / 40,568.55 = 0.8 s prompt_tokens 32768 tokens prefill_tok_s 40568.55 tok/s No published prefill measurement for this hardware; scaled from peak compute (estimate). Model q_dim unknown — attention FLOPs omitted, so long-prompt TTFT is understated.	$100,000 (derived) DERIVEDNVIDIA DGX Station GB300 (Grace Blackwell Ultra), 748GB observed priceArithmetic on observed inputs only. No tunable assumption enters it. observed_price = Σ (cheapest observed unit price × quantity) $100,000 = $100,000 NVIDIA DGX Station GB300 (Grace Blackwell Ultra), 748GB 9999999 ¢ — Central Computers · new · tier 1/jsonld_product Listed price only: shipping and sales tax are NOT in this figure. One price per part — the cheapest monitored, non-quarantined observation.
2× RTX PRO 6000 Blackwell Max-Q Workstation Edition, 96GB · EPYC 9015 · 8× M321R4GA3BB6-CQK 32GB DDR5-4800 RDIMM 429 GB (derived) DERIVED2× RTX PRO 6000 Blackwell Max-Q Workstation Edition, 96GB · EPYC 9015 · 8× M321R4GA3BB6-CQK 32GB DDR5-4800 RDIMM usable memoryArithmetic on observed inputs only. No tunable assumption enters it. available_gb = Σ usable_gb over memory pools GPU VRAM 186.24 + System RAM 243.2 = 429.44 GB GPU VRAM.usable_gb 186.24 GB — 192 GB physical System RAM.usable_gb 243.2 GB — 256 GB physical Usable memory is below physical memory for at least one pool (OS/driver reservation; editable in Settings). usable 1,792 (observed) OBSERVED2× RTX PRO 6000 Blackwell Max-Q Workstation Edition, 96GB · EPYC 9015 · 8× M321R4GA3BB6-CQK 32GB DDR5-4800 RDIMM peak memory bandwidthMeasured or published, passed through unchanged — not computed here. peak memory bandwidth — published at the source, passed through unchanged; nothing is computed here 1,792 GB/s, as recorded from the GPU VRAM spec of 2× RTX PRO 6000 Blackwell Max-Q Workstation Edition, 96GB · EPYC 9015 · 8× M321R4GA3BB6-CQK 32GB DDR5-4800 RDIMM peak_memory_bandwidth 1,792 GB/s — the GPU VRAM spec of 2× RTX PRO 6000 Blackwell Max-Q Workstation Edition, 96GB · EPYC 9015 · 8× M321R4GA3BB6-CQK 32GB DDR5-4800 RDIMM Theoretical peak. Decode reaches a fraction of it — the efficiency factor is in the decode speed's own workings. The fastest pool. A machine with more than one pool places weights fastest-pool-first. Observed, not derived: this app records the figure and its source, and does not re-measure it. GB/s	Fast	43.4 tok/s (estimate) ESTIMATEQwen3-VL 235B-A22B Instruct decode speed on 2× RTX PRO 6000 Blackwell Max-Q Workstation Edition, 96GB · EPYC 9015 · 8× M321R4GA3BB6-CQK 32GB DDR5-4800 RDIMMArithmetic that also depends on a tunable assumption, listed below. tok_s = 1 / (bytes_per_token / (bandwidth × efficiency) + overhead) GPU VRAM: 18.474 GB ÷ (1,792 GB/s × 0.48) + 1.58 ms fixed per-token overhead → 43.37 tok/s GPU VRAM.bandwidth_gbs 1792 GB/s — component spec GPU VRAM.efficiency_moe 0.48 — Settings efficiency.gpu_vram — Discrete GPU: 10 dense + 3 MoE points; fit on published benchmarks, leave-one-out validated — docs/research/12b-estimator-validation.md GPU VRAM.active_share 1 — placement per_token_overhead_ms 1.58 ms — Settings efficiency.gpu_vram.overheadMs — calibrated fit (D-013): Discrete GPU: 10 dense + 3 MoE points; fit on published benchmarks, leave-one-out validated — docs/research/12b-estimator-validation.md bytes_per_token_gb 18.4738 GB MoE: only active parameters are read per token (SPEC §3). Effective bits include embeddings and higher-precision tensors in mixed quants. Assumes active weights are quantized like the file average. KV cache is re-read on every token (DECISIONS D-008). layer split across 2 GPUs: capacity adds, per-token bandwidth does not (layers run sequentially) Per-token overhead of System RAM (44.61 ms) is not charged: this model does not read from that pool. Batch size 1 (single user, interactive).	Unavailable — No prefill measurement or compute rating for DIY: 2× NVIDIA RTX PRO 6000 Blackwell Max-Q Workstation Edition, 96GB + AMD EPYC 9015. (missing: measured prefill_tokens_sec, compute_tflops_fp16)	$27,923 (derived) DERIVED2× RTX PRO 6000 Blackwell Max-Q Workstation Edition, 96GB · EPYC 9015 · 8× M321R4GA3BB6-CQK 32GB DDR5-4800 RDIMM observed priceArithmetic on observed inputs only. No tunable assumption enters it. observed_price = Σ (cheapest observed unit price × quantity) $598 + $839 + $214 × 8 + $11,931 × 2 + $70 + $180 + $660 = $27,923 AMD EPYC 9015 59844 ¢ — SabrePC · new · tier 2/sabrepc_ngstate ASRock Rack GENOAD8QM3-2T/BCM 83900 ¢ — Newegg · new · tier 3/newegg_realtime Samsung M321R4GA3BB6-CQK 32GB DDR5-4800 RDIMM 21427 ¢ — SabrePC · new · tier 2/sabrepc_ngstate NVIDIA RTX PRO 6000 Blackwell Max-Q Workstation Edition, 96GB 1193061 ¢ — SabrePC · new · tier 2/sabrepc_ngstate ARCTIC Freezer 4U-SP5 6999 ¢ — ARCTIC · new · tier 1/arctic_jsonld Phanteks Enthoo Pro 2 Server Edition 17999 ¢ — Newegg · new · tier 3/newegg_realtime Seasonic PRIME PX-2200 65999 ¢ — Newegg · new · tier 3/newegg_realtime Listed price only: shipping and sales tax are NOT in this figure. One price per part — the cheapest monitored, non-quarantined observation.
3× RTX PRO 6000 Blackwell Max-Q Workstation Edition, 96GB · EPYC 9015 · 8× M321R4GA3BB6-CQK 32GB DDR5-4800 RDIMM 523 GB (derived) DERIVED3× RTX PRO 6000 Blackwell Max-Q Workstation Edition, 96GB · EPYC 9015 · 8× M321R4GA3BB6-CQK 32GB DDR5-4800 RDIMM usable memoryArithmetic on observed inputs only. No tunable assumption enters it. available_gb = Σ usable_gb over memory pools GPU VRAM 279.36 + System RAM 243.2 = 522.56 GB GPU VRAM.usable_gb 279.36 GB — 288 GB physical System RAM.usable_gb 243.2 GB — 256 GB physical Usable memory is below physical memory for at least one pool (OS/driver reservation; editable in Settings). usable 1,792 (observed) OBSERVED3× RTX PRO 6000 Blackwell Max-Q Workstation Edition, 96GB · EPYC 9015 · 8× M321R4GA3BB6-CQK 32GB DDR5-4800 RDIMM peak memory bandwidthMeasured or published, passed through unchanged — not computed here. peak memory bandwidth — published at the source, passed through unchanged; nothing is computed here 1,792 GB/s, as recorded from the GPU VRAM spec of 3× RTX PRO 6000 Blackwell Max-Q Workstation Edition, 96GB · EPYC 9015 · 8× M321R4GA3BB6-CQK 32GB DDR5-4800 RDIMM peak_memory_bandwidth 1,792 GB/s — the GPU VRAM spec of 3× RTX PRO 6000 Blackwell Max-Q Workstation Edition, 96GB · EPYC 9015 · 8× M321R4GA3BB6-CQK 32GB DDR5-4800 RDIMM Theoretical peak. Decode reaches a fraction of it — the efficiency factor is in the decode speed's own workings. The fastest pool. A machine with more than one pool places weights fastest-pool-first. Observed, not derived: this app records the figure and its source, and does not re-measure it. GB/s	Fast	43.4 tok/s (estimate) ESTIMATEQwen3-VL 235B-A22B Instruct decode speed on 3× RTX PRO 6000 Blackwell Max-Q Workstation Edition, 96GB · EPYC 9015 · 8× M321R4GA3BB6-CQK 32GB DDR5-4800 RDIMMArithmetic that also depends on a tunable assumption, listed below. tok_s = 1 / (bytes_per_token / (bandwidth × efficiency) + overhead) GPU VRAM: 18.474 GB ÷ (1,792 GB/s × 0.48) + 1.58 ms fixed per-token overhead → 43.37 tok/s GPU VRAM.bandwidth_gbs 1792 GB/s — component spec GPU VRAM.efficiency_moe 0.48 — Settings efficiency.gpu_vram — Discrete GPU: 10 dense + 3 MoE points; fit on published benchmarks, leave-one-out validated — docs/research/12b-estimator-validation.md GPU VRAM.active_share 1 — placement per_token_overhead_ms 1.58 ms — Settings efficiency.gpu_vram.overheadMs — calibrated fit (D-013): Discrete GPU: 10 dense + 3 MoE points; fit on published benchmarks, leave-one-out validated — docs/research/12b-estimator-validation.md bytes_per_token_gb 18.4738 GB MoE: only active parameters are read per token (SPEC §3). Effective bits include embeddings and higher-precision tensors in mixed quants. Assumes active weights are quantized like the file average. KV cache is re-read on every token (DECISIONS D-008). layer split across 3 GPUs: capacity adds, per-token bandwidth does not (layers run sequentially) Per-token overhead of System RAM (44.61 ms) is not charged: this model does not read from that pool. Batch size 1 (single user, interactive).	Unavailable — No prefill measurement or compute rating for DIY: 3× NVIDIA RTX PRO 6000 Blackwell Max-Q Workstation Edition, 96GB + AMD EPYC 9015. (missing: measured prefill_tokens_sec, compute_tflops_fp16)	$39,853 (derived) DERIVED3× RTX PRO 6000 Blackwell Max-Q Workstation Edition, 96GB · EPYC 9015 · 8× M321R4GA3BB6-CQK 32GB DDR5-4800 RDIMM observed priceArithmetic on observed inputs only. No tunable assumption enters it. observed_price = Σ (cheapest observed unit price × quantity) $598 + $839 + $214 × 8 + $11,931 × 3 + $70 + $180 + $660 = $39,853 AMD EPYC 9015 59844 ¢ — SabrePC · new · tier 2/sabrepc_ngstate ASRock Rack GENOAD8QM3-2T/BCM 83900 ¢ — Newegg · new · tier 3/newegg_realtime Samsung M321R4GA3BB6-CQK 32GB DDR5-4800 RDIMM 21427 ¢ — SabrePC · new · tier 2/sabrepc_ngstate NVIDIA RTX PRO 6000 Blackwell Max-Q Workstation Edition, 96GB 1193061 ¢ — SabrePC · new · tier 2/sabrepc_ngstate ARCTIC Freezer 4U-SP5 6999 ¢ — ARCTIC · new · tier 1/arctic_jsonld Phanteks Enthoo Pro 2 Server Edition 17999 ¢ — Newegg · new · tier 3/newegg_realtime Seasonic PRIME PX-2200 65999 ¢ — Newegg · new · tier 3/newegg_realtime Listed price only: shipping and sales tax are NOT in this figure. One price per part — the cheapest monitored, non-quarantined observation.
Apple Mac Studio (2026), M5 Ultra 30-core CPU / 64-core GPU, 256GB 192 GB (derived) DERIVEDApple Mac Studio (2026), M5 Ultra 30-core CPU / 64-core GPU, 256GB usable memoryArithmetic on observed inputs only. No tunable assumption enters it. available_gb = Σ usable_gb over memory pools unified memory 192 = 192 GB unified memory.usable_gb 192 GB — 256 GB physical Usable memory is below physical memory for at least one pool (OS/driver reservation; editable in Settings). usable 1,200 (observed) OBSERVEDApple Mac Studio (2026), M5 Ultra 30-core CPU / 64-core GPU, 256GB peak memory bandwidthMeasured or published, passed through unchanged — not computed here. peak memory bandwidth — published at the source, passed through unchanged; nothing is computed here 1,200 GB/s, as recorded from the unified memory spec of Apple Mac Studio (2026), M5 Ultra 30-core CPU / 64-core GPU, 256GB peak_memory_bandwidth 1,200 GB/s — the unified memory spec of Apple Mac Studio (2026), M5 Ultra 30-core CPU / 64-core GPU, 256GB Theoretical peak. Decode reaches a fraction of it — the efficiency factor is in the decode speed's own workings. The fastest pool. A machine with more than one pool places weights fastest-pool-first. Observed, not derived: this app records the figure and its source, and does not re-measure it. GB/s	Interactive	28.8 tok/s (estimate) ESTIMATEQwen3-VL 235B-A22B Instruct decode speed on Apple Mac Studio (2026), M5 Ultra 30-core CPU / 64-core GPU, 256GBArithmetic that also depends on a tunable assumption, listed below. tok_s = 1 / (bytes_per_token / (bandwidth × efficiency) + overhead) unified memory: 18.474 GB ÷ (1,200 GB/s × 0.48) + 2.71 ms fixed per-token overhead → 28.75 tok/s unified memory.bandwidth_gbs 1200 GB/s — component spec unified memory.efficiency_moe 0.48 — Settings efficiency.apple_unified — Apple unified: 17 dense + 6 MoE points; fit on published benchmarks, leave-one-out validated — docs/research/12b-estimator-validation.md unified memory.active_share 1 — placement per_token_overhead_ms 2.71 ms — Settings efficiency.apple_unified.overheadMs — calibrated fit (D-013): Apple unified: 17 dense + 6 MoE points; fit on published benchmarks, leave-one-out validated — docs/research/12b-estimator-validation.md bytes_per_token_gb 18.4738 GB MoE: only active parameters are read per token (SPEC §3). Effective bits include embeddings and higher-precision tensors in mixed quants. Assumes active weights are quantized like the file average. KV cache is re-read on every token (DECISIONS D-008). Batch size 1 (single user, interactive).	Unavailable — No prefill measurement or compute rating for Apple Mac Studio (2026), M5 Ultra 30-core CPU / 64-core GPU, 256GB. (missing: measured prefill_tokens_sec, compute_tflops_fp16)	$9,499 (derived) DERIVEDApple Mac Studio (2026), M5 Ultra 30-core CPU / 64-core GPU, 256GB observed priceArithmetic on observed inputs only. No tunable assumption enters it. observed_price = Σ (cheapest observed unit price × quantity) $9,499 = $9,499 Apple Mac Studio (2026), M5 Ultra 30-core CPU / 64-core GPU, 256GB 949900 ¢ — Apple · new · tier 3/apple_store Listed price only: shipping and sales tax are NOT in this figure. One price per part — the cheapest monitored, non-quarantined observation.
Apple Mac Studio (2026), M5 Ultra 36-core CPU / 80-core GPU, 256GB 192 GB (derived) DERIVEDApple Mac Studio (2026), M5 Ultra 36-core CPU / 80-core GPU, 256GB usable memoryArithmetic on observed inputs only. No tunable assumption enters it. available_gb = Σ usable_gb over memory pools unified memory 192 = 192 GB unified memory.usable_gb 192 GB — 256 GB physical Usable memory is below physical memory for at least one pool (OS/driver reservation; editable in Settings). usable 1,200 (observed) OBSERVEDApple Mac Studio (2026), M5 Ultra 36-core CPU / 80-core GPU, 256GB peak memory bandwidthMeasured or published, passed through unchanged — not computed here. peak memory bandwidth — published at the source, passed through unchanged; nothing is computed here 1,200 GB/s, as recorded from the unified memory spec of Apple Mac Studio (2026), M5 Ultra 36-core CPU / 80-core GPU, 256GB peak_memory_bandwidth 1,200 GB/s — the unified memory spec of Apple Mac Studio (2026), M5 Ultra 36-core CPU / 80-core GPU, 256GB Theoretical peak. Decode reaches a fraction of it — the efficiency factor is in the decode speed's own workings. The fastest pool. A machine with more than one pool places weights fastest-pool-first. Observed, not derived: this app records the figure and its source, and does not re-measure it. GB/s	Interactive	28.8 tok/s (estimate) ESTIMATEQwen3-VL 235B-A22B Instruct decode speed on Apple Mac Studio (2026), M5 Ultra 36-core CPU / 80-core GPU, 256GBArithmetic that also depends on a tunable assumption, listed below. tok_s = 1 / (bytes_per_token / (bandwidth × efficiency) + overhead) unified memory: 18.474 GB ÷ (1,200 GB/s × 0.48) + 2.71 ms fixed per-token overhead → 28.75 tok/s unified memory.bandwidth_gbs 1200 GB/s — component spec unified memory.efficiency_moe 0.48 — Settings efficiency.apple_unified — Apple unified: 17 dense + 6 MoE points; fit on published benchmarks, leave-one-out validated — docs/research/12b-estimator-validation.md unified memory.active_share 1 — placement per_token_overhead_ms 2.71 ms — Settings efficiency.apple_unified.overheadMs — calibrated fit (D-013): Apple unified: 17 dense + 6 MoE points; fit on published benchmarks, leave-one-out validated — docs/research/12b-estimator-validation.md bytes_per_token_gb 18.4738 GB MoE: only active parameters are read per token (SPEC §3). Effective bits include embeddings and higher-precision tensors in mixed quants. Assumes active weights are quantized like the file average. KV cache is re-read on every token (DECISIONS D-008). Batch size 1 (single user, interactive).	21.2 s (estimate) ESTIMATEQwen3-VL 235B-A22B Instruct time to first token on Apple Mac Studio (2026), M5 Ultra 36-core CPU / 80-core GPU, 256GBArithmetic that also depends on a tunable assumption, listed below. ttft_s = prompt_tokens / prefill_tok_s = 32,768 / 1,545.18 = 21.2 s prompt_tokens 32768 tokens prefill_tok_s 1545.18 tok/s Scaled from a published measurement by active parameters (prefill FLOPs ∝ active params). Attention cost difference between the reference prompt and this prompt is not rescaled.	$10,799 (derived) DERIVEDApple Mac Studio (2026), M5 Ultra 36-core CPU / 80-core GPU, 256GB observed priceArithmetic on observed inputs only. No tunable assumption enters it. observed_price = Σ (cheapest observed unit price × quantity) $10,799 = $10,799 Apple Mac Studio (2026), M5 Ultra 36-core CPU / 80-core GPU, 256GB 1079900 ¢ — Apple · new · tier 3/apple_store Listed price only: shipping and sales tax are NOT in this figure. One price per part — the cheapest monitored, non-quarantined observation.
Apple Mac Studio (2026), M5 Ultra 36-core CPU / 80-core GPU, 512GB 384 GB (derived) DERIVEDApple Mac Studio (2026), M5 Ultra 36-core CPU / 80-core GPU, 512GB usable memoryArithmetic on observed inputs only. No tunable assumption enters it. available_gb = Σ usable_gb over memory pools unified memory 384 = 384 GB unified memory.usable_gb 384 GB — 512 GB physical Usable memory is below physical memory for at least one pool (OS/driver reservation; editable in Settings). usable 1,200 (observed) OBSERVEDApple Mac Studio (2026), M5 Ultra 36-core CPU / 80-core GPU, 512GB peak memory bandwidthMeasured or published, passed through unchanged — not computed here. peak memory bandwidth — published at the source, passed through unchanged; nothing is computed here 1,200 GB/s, as recorded from the unified memory spec of Apple Mac Studio (2026), M5 Ultra 36-core CPU / 80-core GPU, 512GB peak_memory_bandwidth 1,200 GB/s — the unified memory spec of Apple Mac Studio (2026), M5 Ultra 36-core CPU / 80-core GPU, 512GB Theoretical peak. Decode reaches a fraction of it — the efficiency factor is in the decode speed's own workings. The fastest pool. A machine with more than one pool places weights fastest-pool-first. Observed, not derived: this app records the figure and its source, and does not re-measure it. GB/s	Interactive	28.8 tok/s (estimate) ESTIMATEQwen3-VL 235B-A22B Instruct decode speed on Apple Mac Studio (2026), M5 Ultra 36-core CPU / 80-core GPU, 512GBArithmetic that also depends on a tunable assumption, listed below. tok_s = 1 / (bytes_per_token / (bandwidth × efficiency) + overhead) unified memory: 18.474 GB ÷ (1,200 GB/s × 0.48) + 2.71 ms fixed per-token overhead → 28.75 tok/s unified memory.bandwidth_gbs 1200 GB/s — component spec unified memory.efficiency_moe 0.48 — Settings efficiency.apple_unified — Apple unified: 17 dense + 6 MoE points; fit on published benchmarks, leave-one-out validated — docs/research/12b-estimator-validation.md unified memory.active_share 1 — placement per_token_overhead_ms 2.71 ms — Settings efficiency.apple_unified.overheadMs — calibrated fit (D-013): Apple unified: 17 dense + 6 MoE points; fit on published benchmarks, leave-one-out validated — docs/research/12b-estimator-validation.md bytes_per_token_gb 18.4738 GB MoE: only active parameters are read per token (SPEC §3). Effective bits include embeddings and higher-precision tensors in mixed quants. Assumes active weights are quantized like the file average. KV cache is re-read on every token (DECISIONS D-008). Batch size 1 (single user, interactive).	21.2 s (estimate) ESTIMATEQwen3-VL 235B-A22B Instruct time to first token on Apple Mac Studio (2026), M5 Ultra 36-core CPU / 80-core GPU, 512GBArithmetic that also depends on a tunable assumption, listed below. ttft_s = prompt_tokens / prefill_tok_s = 32,768 / 1,545.18 = 21.2 s prompt_tokens 32768 tokens prefill_tok_s 1545.18 tok/s Scaled from a published measurement by active parameters (prefill FLOPs ∝ active params). Attention cost difference between the reference prompt and this prompt is not rescaled.	Unavailable — No monitored retailer has published a price for this machine: Apple Mac Studio (2026), M5 Ultra 36-core CPU / 80-core GPU, 512GB. (missing: Apple Mac Studio (2026), M5 Ultra 36-core CPU / 80-core GPU, 512GB)
1× RTX PRO 6000 Blackwell Workstation Edition, 96GB · EPYC 9015 · 8× M321R4GA3BB6-CQK 32GB DDR5-4800 RDIMM 336 GB (derived) DERIVED1× RTX PRO 6000 Blackwell Workstation Edition, 96GB · EPYC 9015 · 8× M321R4GA3BB6-CQK 32GB DDR5-4800 RDIMM usable memoryArithmetic on observed inputs only. No tunable assumption enters it. available_gb = Σ usable_gb over memory pools GPU VRAM 93.12 + System RAM 243.2 = 336.32 GB GPU VRAM.usable_gb 93.12 GB — 96 GB physical System RAM.usable_gb 243.2 GB — 256 GB physical Usable memory is below physical memory for at least one pool (OS/driver reservation; editable in Settings). usable 1,792 (observed) OBSERVED1× RTX PRO 6000 Blackwell Workstation Edition, 96GB · EPYC 9015 · 8× M321R4GA3BB6-CQK 32GB DDR5-4800 RDIMM peak memory bandwidthMeasured or published, passed through unchanged — not computed here. peak memory bandwidth — published at the source, passed through unchanged; nothing is computed here 1,792 GB/s, as recorded from the GPU VRAM spec of 1× RTX PRO 6000 Blackwell Workstation Edition, 96GB · EPYC 9015 · 8× M321R4GA3BB6-CQK 32GB DDR5-4800 RDIMM peak_memory_bandwidth 1,792 GB/s — the GPU VRAM spec of 1× RTX PRO 6000 Blackwell Workstation Edition, 96GB · EPYC 9015 · 8× M321R4GA3BB6-CQK 32GB DDR5-4800 RDIMM Theoretical peak. Decode reaches a fraction of it — the efficiency factor is in the decode speed's own workings. The fastest pool. A machine with more than one pool places weights fastest-pool-first. Observed, not derived: this app records the figure and its source, and does not re-measure it. GB/s	Interactive	15.7 tok/s (estimate) ESTIMATEQwen3-VL 235B-A22B Instruct decode speed on 1× RTX PRO 6000 Blackwell Workstation Edition, 96GB · EPYC 9015 · 8× M321R4GA3BB6-CQK 32GB DDR5-4800 RDIMMArithmetic that also depends on a tunable assumption, listed below. tok_s = 1 / (Σ_pool bytes_per_token × share_pool / (bandwidth_pool × efficiency_pool) + Σ_pool share_pool × overhead_pool) GPU VRAM: 11.505 GB ÷ (1,792 GB/s × 0.48) + System RAM: 6.968 GB ÷ (307.2 GB/s × 0.7) + 17.81 ms fixed per-token overhead (share-weighted) → 15.73 tok/s GPU VRAM.bandwidth_gbs 1792 GB/s — component spec GPU VRAM.efficiency_moe 0.48 — Settings efficiency.gpu_vram — Discrete GPU: 10 dense + 3 MoE points; fit on published benchmarks, leave-one-out validated — docs/research/12b-estimator-validation.md GPU VRAM.active_share 0.6228 — placement System RAM.bandwidth_gbs 307.2 GB/s — component spec System RAM.efficiency_moe 0.7 — Settings efficiency.server_ddr5 — Server DDR5/MRDIMM single socket: 6 dense + 7 MoE points (small n); fit on published benchmarks, leave-one-out validated — docs/research/12b-estimator-validation.md. The 44.61 ms per-token overhead is a CPU-engine constant, not a property of DDR5: its fit is set by a 128-core Xeon 6980P running llama.cpp (55.3 ms/token on a 3B model) and SGLang (17.0 ms) with the CPU executing every layer (research 12 §6). A GPU spilling into this memory pays it only for the share of each token read from RAM (lib/perf/decode.ts refinement 3). System RAM.active_share 0.3772 — placement per_token_overhead_ms 17.811206136367083 ms — derived: 0.623 × 1.58 ms (GPU VRAM) + 0.377 × 44.61 ms (System RAM) = 17.81 ms bytes_per_token_gb 18.4738 GB MoE: only active parameters are read per token (SPEC §3). Effective bits include embeddings and higher-precision tensors in mixed quants. Assumes active weights are quantized like the file average. KV cache is re-read on every token (DECISIONS D-008). Pools are read sequentially per token, so per-pool times add (harmonic combination). Per-token overhead on a split placement is DERIVED, not measured: each pool's own single-pool figure is charged in proportion to the share of each token it serves, because the fixed cost is per operation and the split puts that share of the operations on each device. Calibrated with that pool serving the whole model (D-013): GPU VRAM (1.58 ms), System RAM (44.61 ms). Batch size 1 (single user, interactive).	Unavailable — No prefill measurement or compute rating for DIY: NVIDIA RTX PRO 6000 Blackwell Workstation Edition, 96GB + AMD EPYC 9015. (missing: measured prefill_tokens_sec, compute_tflops_fp16)	$14,857 (derived) DERIVED1× RTX PRO 6000 Blackwell Workstation Edition, 96GB · EPYC 9015 · 8× M321R4GA3BB6-CQK 32GB DDR5-4800 RDIMM observed priceArithmetic on observed inputs only. No tunable assumption enters it. observed_price = Σ (cheapest observed unit price × quantity) $598 + $839 + $214 × 8 + $10,795 + $70 + $180 + $660 = $14,857 AMD EPYC 9015 59844 ¢ — SabrePC · new · tier 2/sabrepc_ngstate ASRock Rack GENOAD8QM3-2T/BCM 83900 ¢ — Newegg · new · tier 3/newegg_realtime Samsung M321R4GA3BB6-CQK 32GB DDR5-4800 RDIMM 21427 ¢ — SabrePC · new · tier 2/sabrepc_ngstate NVIDIA RTX PRO 6000 Blackwell Workstation Edition, 96GB 1079547 ¢ — SabrePC · new · tier 2/sabrepc_ngstate ARCTIC Freezer 4U-SP5 6999 ¢ — ARCTIC · new · tier 1/arctic_jsonld Phanteks Enthoo Pro 2 Server Edition 17999 ¢ — Newegg · new · tier 3/newegg_realtime Seasonic PRIME PX-2200 65999 ¢ — Newegg · new · tier 3/newegg_realtime Listed price only: shipping and sales tax are NOT in this figure. One price per part — the cheapest monitored, non-quarantined observation.
1× RTX PRO 6000 Blackwell Max-Q Workstation Edition, 96GB · EPYC 9015 · 8× M321R4GA3BB6-CQK 32GB DDR5-4800 RDIMM 336 GB (derived) DERIVED1× RTX PRO 6000 Blackwell Max-Q Workstation Edition, 96GB · EPYC 9015 · 8× M321R4GA3BB6-CQK 32GB DDR5-4800 RDIMM usable memoryArithmetic on observed inputs only. No tunable assumption enters it. available_gb = Σ usable_gb over memory pools GPU VRAM 93.12 + System RAM 243.2 = 336.32 GB GPU VRAM.usable_gb 93.12 GB — 96 GB physical System RAM.usable_gb 243.2 GB — 256 GB physical Usable memory is below physical memory for at least one pool (OS/driver reservation; editable in Settings). usable 1,792 (observed) OBSERVED1× RTX PRO 6000 Blackwell Max-Q Workstation Edition, 96GB · EPYC 9015 · 8× M321R4GA3BB6-CQK 32GB DDR5-4800 RDIMM peak memory bandwidthMeasured or published, passed through unchanged — not computed here. peak memory bandwidth — published at the source, passed through unchanged; nothing is computed here 1,792 GB/s, as recorded from the GPU VRAM spec of 1× RTX PRO 6000 Blackwell Max-Q Workstation Edition, 96GB · EPYC 9015 · 8× M321R4GA3BB6-CQK 32GB DDR5-4800 RDIMM peak_memory_bandwidth 1,792 GB/s — the GPU VRAM spec of 1× RTX PRO 6000 Blackwell Max-Q Workstation Edition, 96GB · EPYC 9015 · 8× M321R4GA3BB6-CQK 32GB DDR5-4800 RDIMM Theoretical peak. Decode reaches a fraction of it — the efficiency factor is in the decode speed's own workings. The fastest pool. A machine with more than one pool places weights fastest-pool-first. Observed, not derived: this app records the figure and its source, and does not re-measure it. GB/s	Interactive	15.7 tok/s (estimate) ESTIMATEQwen3-VL 235B-A22B Instruct decode speed on 1× RTX PRO 6000 Blackwell Max-Q Workstation Edition, 96GB · EPYC 9015 · 8× M321R4GA3BB6-CQK 32GB DDR5-4800 RDIMMArithmetic that also depends on a tunable assumption, listed below. tok_s = 1 / (Σ_pool bytes_per_token × share_pool / (bandwidth_pool × efficiency_pool) + Σ_pool share_pool × overhead_pool) GPU VRAM: 11.505 GB ÷ (1,792 GB/s × 0.48) + System RAM: 6.968 GB ÷ (307.2 GB/s × 0.7) + 17.81 ms fixed per-token overhead (share-weighted) → 15.73 tok/s GPU VRAM.bandwidth_gbs 1792 GB/s — component spec GPU VRAM.efficiency_moe 0.48 — Settings efficiency.gpu_vram — Discrete GPU: 10 dense + 3 MoE points; fit on published benchmarks, leave-one-out validated — docs/research/12b-estimator-validation.md GPU VRAM.active_share 0.6228 — placement System RAM.bandwidth_gbs 307.2 GB/s — component spec System RAM.efficiency_moe 0.7 — Settings efficiency.server_ddr5 — Server DDR5/MRDIMM single socket: 6 dense + 7 MoE points (small n); fit on published benchmarks, leave-one-out validated — docs/research/12b-estimator-validation.md. The 44.61 ms per-token overhead is a CPU-engine constant, not a property of DDR5: its fit is set by a 128-core Xeon 6980P running llama.cpp (55.3 ms/token on a 3B model) and SGLang (17.0 ms) with the CPU executing every layer (research 12 §6). A GPU spilling into this memory pays it only for the share of each token read from RAM (lib/perf/decode.ts refinement 3). System RAM.active_share 0.3772 — placement per_token_overhead_ms 17.811206136367083 ms — derived: 0.623 × 1.58 ms (GPU VRAM) + 0.377 × 44.61 ms (System RAM) = 17.81 ms bytes_per_token_gb 18.4738 GB MoE: only active parameters are read per token (SPEC §3). Effective bits include embeddings and higher-precision tensors in mixed quants. Assumes active weights are quantized like the file average. KV cache is re-read on every token (DECISIONS D-008). Pools are read sequentially per token, so per-pool times add (harmonic combination). Per-token overhead on a split placement is DERIVED, not measured: each pool's own single-pool figure is charged in proportion to the share of each token it serves, because the fixed cost is per operation and the split puts that share of the operations on each device. Calibrated with that pool serving the whole model (D-013): GPU VRAM (1.58 ms), System RAM (44.61 ms). Batch size 1 (single user, interactive).	Unavailable — No prefill measurement or compute rating for DIY: NVIDIA RTX PRO 6000 Blackwell Max-Q Workstation Edition, 96GB + AMD EPYC 9015. (missing: measured prefill_tokens_sec, compute_tflops_fp16)	$15,992 (derived) DERIVED1× RTX PRO 6000 Blackwell Max-Q Workstation Edition, 96GB · EPYC 9015 · 8× M321R4GA3BB6-CQK 32GB DDR5-4800 RDIMM observed priceArithmetic on observed inputs only. No tunable assumption enters it. observed_price = Σ (cheapest observed unit price × quantity) $598 + $839 + $214 × 8 + $11,931 + $70 + $180 + $660 = $15,992 AMD EPYC 9015 59844 ¢ — SabrePC · new · tier 2/sabrepc_ngstate ASRock Rack GENOAD8QM3-2T/BCM 83900 ¢ — Newegg · new · tier 3/newegg_realtime Samsung M321R4GA3BB6-CQK 32GB DDR5-4800 RDIMM 21427 ¢ — SabrePC · new · tier 2/sabrepc_ngstate NVIDIA RTX PRO 6000 Blackwell Max-Q Workstation Edition, 96GB 1193061 ¢ — SabrePC · new · tier 2/sabrepc_ngstate ARCTIC Freezer 4U-SP5 6999 ¢ — ARCTIC · new · tier 1/arctic_jsonld Phanteks Enthoo Pro 2 Server Edition 17999 ¢ — Newegg · new · tier 3/newegg_realtime Seasonic PRIME PX-2200 65999 ¢ — Newegg · new · tier 3/newegg_realtime Listed price only: shipping and sales tax are NOT in this figure. One price per part — the cheapest monitored, non-quarantined observation.
Hybrid · RTX PRO 6000 Blackwell Workstation Edition, 96GB + EPYC 9015 · 8× M321R8GA0BB0-CQK 64GB DDR5-4800 RDIMM 580 GB (derived) DERIVEDHybrid · RTX PRO 6000 Blackwell Workstation Edition, 96GB + EPYC 9015 · 8× M321R8GA0BB0-CQK 64GB DDR5-4800 RDIMM usable memoryArithmetic on observed inputs only. No tunable assumption enters it. available_gb = Σ usable_gb over memory pools GPU VRAM 93.12 + System RAM 486.4 = 579.52 GB GPU VRAM.usable_gb 93.12 GB — 96 GB physical System RAM.usable_gb 486.4 GB — 512 GB physical Usable memory is below physical memory for at least one pool (OS/driver reservation; editable in Settings). usable 1,792 (observed) OBSERVEDHybrid · RTX PRO 6000 Blackwell Workstation Edition, 96GB + EPYC 9015 · 8× M321R8GA0BB0-CQK 64GB DDR5-4800 RDIMM peak memory bandwidthMeasured or published, passed through unchanged — not computed here. peak memory bandwidth — published at the source, passed through unchanged; nothing is computed here 1,792 GB/s, as recorded from the GPU VRAM spec of Hybrid · RTX PRO 6000 Blackwell Workstation Edition, 96GB + EPYC 9015 · 8× M321R8GA0BB0-CQK 64GB DDR5-4800 RDIMM peak_memory_bandwidth 1,792 GB/s — the GPU VRAM spec of Hybrid · RTX PRO 6000 Blackwell Workstation Edition, 96GB + EPYC 9015 · 8× M321R8GA0BB0-CQK 64GB DDR5-4800 RDIMM Theoretical peak. Decode reaches a fraction of it — the efficiency factor is in the decode speed's own workings. The fastest pool. A machine with more than one pool places weights fastest-pool-first. Observed, not derived: this app records the figure and its source, and does not re-measure it. GB/s	Interactive	15.7 tok/s (estimate) ESTIMATEQwen3-VL 235B-A22B Instruct decode speed on Hybrid · RTX PRO 6000 Blackwell Workstation Edition, 96GB + EPYC 9015 · 8× M321R8GA0BB0-CQK 64GB DDR5-4800 RDIMMArithmetic that also depends on a tunable assumption, listed below. tok_s = 1 / (Σ_pool bytes_per_token × share_pool / (bandwidth_pool × efficiency_pool) + Σ_pool share_pool × overhead_pool) GPU VRAM: 11.505 GB ÷ (1,792 GB/s × 0.48) + System RAM: 6.968 GB ÷ (307.2 GB/s × 0.7) + 17.81 ms fixed per-token overhead (share-weighted) → 15.73 tok/s GPU VRAM.bandwidth_gbs 1792 GB/s — component spec GPU VRAM.efficiency_moe 0.48 — Settings efficiency.gpu_vram — Discrete GPU: 10 dense + 3 MoE points; fit on published benchmarks, leave-one-out validated — docs/research/12b-estimator-validation.md GPU VRAM.active_share 0.6228 — placement System RAM.bandwidth_gbs 307.2 GB/s — component spec System RAM.efficiency_moe 0.7 — Settings efficiency.server_ddr5 — Server DDR5/MRDIMM single socket: 6 dense + 7 MoE points (small n); fit on published benchmarks, leave-one-out validated — docs/research/12b-estimator-validation.md. The 44.61 ms per-token overhead is a CPU-engine constant, not a property of DDR5: its fit is set by a 128-core Xeon 6980P running llama.cpp (55.3 ms/token on a 3B model) and SGLang (17.0 ms) with the CPU executing every layer (research 12 §6). A GPU spilling into this memory pays it only for the share of each token read from RAM (lib/perf/decode.ts refinement 3). System RAM.active_share 0.3772 — placement per_token_overhead_ms 17.811206136367083 ms — derived: 0.623 × 1.58 ms (GPU VRAM) + 0.377 × 44.61 ms (System RAM) = 17.81 ms bytes_per_token_gb 18.4738 GB MoE: only active parameters are read per token (SPEC §3). Effective bits include embeddings and higher-precision tensors in mixed quants. Assumes active weights are quantized like the file average. KV cache is re-read on every token (DECISIONS D-008). Pools are read sequentially per token, so per-pool times add (harmonic combination). Per-token overhead on a split placement is DERIVED, not measured: each pool's own single-pool figure is charged in proportion to the share of each token it serves, because the fixed cost is per operation and the split puts that share of the operations on each device. Calibrated with that pool serving the whole model (D-013): GPU VRAM (1.58 ms), System RAM (44.61 ms). Batch size 1 (single user, interactive).	Unavailable — No prefill measurement or compute rating for DIY: NVIDIA RTX PRO 6000 Blackwell Workstation Edition, 96GB + AMD EPYC 9015. (missing: measured prefill_tokens_sec, compute_tflops_fp16)	$15,519 (derived) DERIVEDHybrid · RTX PRO 6000 Blackwell Workstation Edition, 96GB + EPYC 9015 · 8× M321R8GA0BB0-CQK 64GB DDR5-4800 RDIMM observed priceArithmetic on observed inputs only. No tunable assumption enters it. observed_price = Σ (cheapest observed unit price × quantity) $598 + $839 + $297 × 8 + $10,795 + $70 + $180 + $660 = $15,519 AMD EPYC 9015 59844 ¢ — SabrePC · new · tier 2/sabrepc_ngstate ASRock Rack GENOAD8QM3-2T/BCM 83900 ¢ — Newegg · new · tier 3/newegg_realtime Samsung M321R8GA0BB0-CQK 64GB DDR5-4800 RDIMM 29707 ¢ — SabrePC · new · tier 2/sabrepc_ngstate NVIDIA RTX PRO 6000 Blackwell Workstation Edition, 96GB 1079547 ¢ — SabrePC · new · tier 2/sabrepc_ngstate ARCTIC Freezer 4U-SP5 6999 ¢ — ARCTIC · new · tier 1/arctic_jsonld Phanteks Enthoo Pro 2 Server Edition 17999 ¢ — Newegg · new · tier 3/newegg_realtime Seasonic PRIME PX-2200 65999 ¢ — Newegg · new · tier 3/newegg_realtime Listed price only: shipping and sales tax are NOT in this figure. One price per part — the cheapest monitored, non-quarantined observation.
Hybrid · RTX PRO 6000 Blackwell Max-Q Workstation Edition, 96GB + EPYC 9015 · 8× M321R8GA0BB0-CQK 64GB DDR5-4800 RDIMM 580 GB (derived) DERIVEDHybrid · RTX PRO 6000 Blackwell Max-Q Workstation Edition, 96GB + EPYC 9015 · 8× M321R8GA0BB0-CQK 64GB DDR5-4800 RDIMM usable memoryArithmetic on observed inputs only. No tunable assumption enters it. available_gb = Σ usable_gb over memory pools GPU VRAM 93.12 + System RAM 486.4 = 579.52 GB GPU VRAM.usable_gb 93.12 GB — 96 GB physical System RAM.usable_gb 486.4 GB — 512 GB physical Usable memory is below physical memory for at least one pool (OS/driver reservation; editable in Settings). usable 1,792 (observed) OBSERVEDHybrid · RTX PRO 6000 Blackwell Max-Q Workstation Edition, 96GB + EPYC 9015 · 8× M321R8GA0BB0-CQK 64GB DDR5-4800 RDIMM peak memory bandwidthMeasured or published, passed through unchanged — not computed here. peak memory bandwidth — published at the source, passed through unchanged; nothing is computed here 1,792 GB/s, as recorded from the GPU VRAM spec of Hybrid · RTX PRO 6000 Blackwell Max-Q Workstation Edition, 96GB + EPYC 9015 · 8× M321R8GA0BB0-CQK 64GB DDR5-4800 RDIMM peak_memory_bandwidth 1,792 GB/s — the GPU VRAM spec of Hybrid · RTX PRO 6000 Blackwell Max-Q Workstation Edition, 96GB + EPYC 9015 · 8× M321R8GA0BB0-CQK 64GB DDR5-4800 RDIMM Theoretical peak. Decode reaches a fraction of it — the efficiency factor is in the decode speed's own workings. The fastest pool. A machine with more than one pool places weights fastest-pool-first. Observed, not derived: this app records the figure and its source, and does not re-measure it. GB/s	Interactive	15.7 tok/s (estimate) ESTIMATEQwen3-VL 235B-A22B Instruct decode speed on Hybrid · RTX PRO 6000 Blackwell Max-Q Workstation Edition, 96GB + EPYC 9015 · 8× M321R8GA0BB0-CQK 64GB DDR5-4800 RDIMMArithmetic that also depends on a tunable assumption, listed below. tok_s = 1 / (Σ_pool bytes_per_token × share_pool / (bandwidth_pool × efficiency_pool) + Σ_pool share_pool × overhead_pool) GPU VRAM: 11.505 GB ÷ (1,792 GB/s × 0.48) + System RAM: 6.968 GB ÷ (307.2 GB/s × 0.7) + 17.81 ms fixed per-token overhead (share-weighted) → 15.73 tok/s GPU VRAM.bandwidth_gbs 1792 GB/s — component spec GPU VRAM.efficiency_moe 0.48 — Settings efficiency.gpu_vram — Discrete GPU: 10 dense + 3 MoE points; fit on published benchmarks, leave-one-out validated — docs/research/12b-estimator-validation.md GPU VRAM.active_share 0.6228 — placement System RAM.bandwidth_gbs 307.2 GB/s — component spec System RAM.efficiency_moe 0.7 — Settings efficiency.server_ddr5 — Server DDR5/MRDIMM single socket: 6 dense + 7 MoE points (small n); fit on published benchmarks, leave-one-out validated — docs/research/12b-estimator-validation.md. The 44.61 ms per-token overhead is a CPU-engine constant, not a property of DDR5: its fit is set by a 128-core Xeon 6980P running llama.cpp (55.3 ms/token on a 3B model) and SGLang (17.0 ms) with the CPU executing every layer (research 12 §6). A GPU spilling into this memory pays it only for the share of each token read from RAM (lib/perf/decode.ts refinement 3). System RAM.active_share 0.3772 — placement per_token_overhead_ms 17.811206136367083 ms — derived: 0.623 × 1.58 ms (GPU VRAM) + 0.377 × 44.61 ms (System RAM) = 17.81 ms bytes_per_token_gb 18.4738 GB MoE: only active parameters are read per token (SPEC §3). Effective bits include embeddings and higher-precision tensors in mixed quants. Assumes active weights are quantized like the file average. KV cache is re-read on every token (DECISIONS D-008). Pools are read sequentially per token, so per-pool times add (harmonic combination). Per-token overhead on a split placement is DERIVED, not measured: each pool's own single-pool figure is charged in proportion to the share of each token it serves, because the fixed cost is per operation and the split puts that share of the operations on each device. Calibrated with that pool serving the whole model (D-013): GPU VRAM (1.58 ms), System RAM (44.61 ms). Batch size 1 (single user, interactive).	Unavailable — No prefill measurement or compute rating for DIY: NVIDIA RTX PRO 6000 Blackwell Max-Q Workstation Edition, 96GB + AMD EPYC 9015. (missing: measured prefill_tokens_sec, compute_tflops_fp16)	$16,655 (derived) DERIVEDHybrid · RTX PRO 6000 Blackwell Max-Q Workstation Edition, 96GB + EPYC 9015 · 8× M321R8GA0BB0-CQK 64GB DDR5-4800 RDIMM observed priceArithmetic on observed inputs only. No tunable assumption enters it. observed_price = Σ (cheapest observed unit price × quantity) $598 + $839 + $297 × 8 + $11,931 + $70 + $180 + $660 = $16,655 AMD EPYC 9015 59844 ¢ — SabrePC · new · tier 2/sabrepc_ngstate ASRock Rack GENOAD8QM3-2T/BCM 83900 ¢ — Newegg · new · tier 3/newegg_realtime Samsung M321R8GA0BB0-CQK 64GB DDR5-4800 RDIMM 29707 ¢ — SabrePC · new · tier 2/sabrepc_ngstate NVIDIA RTX PRO 6000 Blackwell Max-Q Workstation Edition, 96GB 1193061 ¢ — SabrePC · new · tier 2/sabrepc_ngstate ARCTIC Freezer 4U-SP5 6999 ¢ — ARCTIC · new · tier 1/arctic_jsonld Phanteks Enthoo Pro 2 Server Edition 17999 ¢ — Newegg · new · tier 3/newegg_realtime Seasonic PRIME PX-2200 65999 ¢ — Newegg · new · tier 3/newegg_realtime Listed price only: shipping and sales tax are NOT in this figure. One price per part — the cheapest monitored, non-quarantined observation.
Hybrid · RTX PRO 6000 Blackwell Workstation Edition, 96GB + EPYC 9015 · 8× Server Premier KSM64R52BD4-96MC 96GB DDR5-6400 RDIMM 823 GB (derived) DERIVEDHybrid · RTX PRO 6000 Blackwell Workstation Edition, 96GB + EPYC 9015 · 8× Server Premier KSM64R52BD4-96MC 96GB DDR5-6400 RDIMM usable memoryArithmetic on observed inputs only. No tunable assumption enters it. available_gb = Σ usable_gb over memory pools GPU VRAM 93.12 + System RAM 729.6 = 822.72 GB GPU VRAM.usable_gb 93.12 GB — 96 GB physical System RAM.usable_gb 729.5999999999999 GB — 768 GB physical Usable memory is below physical memory for at least one pool (OS/driver reservation; editable in Settings). usable 1,792 (observed) OBSERVEDHybrid · RTX PRO 6000 Blackwell Workstation Edition, 96GB + EPYC 9015 · 8× Server Premier KSM64R52BD4-96MC 96GB DDR5-6400 RDIMM peak memory bandwidthMeasured or published, passed through unchanged — not computed here. peak memory bandwidth — published at the source, passed through unchanged; nothing is computed here 1,792 GB/s, as recorded from the GPU VRAM spec of Hybrid · RTX PRO 6000 Blackwell Workstation Edition, 96GB + EPYC 9015 · 8× Server Premier KSM64R52BD4-96MC 96GB DDR5-6400 RDIMM peak_memory_bandwidth 1,792 GB/s — the GPU VRAM spec of Hybrid · RTX PRO 6000 Blackwell Workstation Edition, 96GB + EPYC 9015 · 8× Server Premier KSM64R52BD4-96MC 96GB DDR5-6400 RDIMM Theoretical peak. Decode reaches a fraction of it — the efficiency factor is in the decode speed's own workings. The fastest pool. A machine with more than one pool places weights fastest-pool-first. Observed, not derived: this app records the figure and its source, and does not re-measure it. GB/s	Interactive	15.7 tok/s (estimate) ESTIMATEQwen3-VL 235B-A22B Instruct decode speed on Hybrid · RTX PRO 6000 Blackwell Workstation Edition, 96GB + EPYC 9015 · 8× Server Premier KSM64R52BD4-96MC 96GB DDR5-6400 RDIMMArithmetic that also depends on a tunable assumption, listed below. tok_s = 1 / (Σ_pool bytes_per_token × share_pool / (bandwidth_pool × efficiency_pool) + Σ_pool share_pool × overhead_pool) GPU VRAM: 11.505 GB ÷ (1,792 GB/s × 0.48) + System RAM: 6.968 GB ÷ (307.2 GB/s × 0.7) + 17.81 ms fixed per-token overhead (share-weighted) → 15.73 tok/s GPU VRAM.bandwidth_gbs 1792 GB/s — component spec GPU VRAM.efficiency_moe 0.48 — Settings efficiency.gpu_vram — Discrete GPU: 10 dense + 3 MoE points; fit on published benchmarks, leave-one-out validated — docs/research/12b-estimator-validation.md GPU VRAM.active_share 0.6228 — placement System RAM.bandwidth_gbs 307.2 GB/s — component spec System RAM.efficiency_moe 0.7 — Settings efficiency.server_ddr5 — Server DDR5/MRDIMM single socket: 6 dense + 7 MoE points (small n); fit on published benchmarks, leave-one-out validated — docs/research/12b-estimator-validation.md. The 44.61 ms per-token overhead is a CPU-engine constant, not a property of DDR5: its fit is set by a 128-core Xeon 6980P running llama.cpp (55.3 ms/token on a 3B model) and SGLang (17.0 ms) with the CPU executing every layer (research 12 §6). A GPU spilling into this memory pays it only for the share of each token read from RAM (lib/perf/decode.ts refinement 3). System RAM.active_share 0.3772 — placement per_token_overhead_ms 17.811206136367083 ms — derived: 0.623 × 1.58 ms (GPU VRAM) + 0.377 × 44.61 ms (System RAM) = 17.81 ms bytes_per_token_gb 18.4738 GB MoE: only active parameters are read per token (SPEC §3). Effective bits include embeddings and higher-precision tensors in mixed quants. Assumes active weights are quantized like the file average. KV cache is re-read on every token (DECISIONS D-008). Pools are read sequentially per token, so per-pool times add (harmonic combination). Per-token overhead on a split placement is DERIVED, not measured: each pool's own single-pool figure is charged in proportion to the share of each token it serves, because the fixed cost is per operation and the split puts that share of the operations on each device. Calibrated with that pool serving the whole model (D-013): GPU VRAM (1.58 ms), System RAM (44.61 ms). Batch size 1 (single user, interactive).	Unavailable — No prefill measurement or compute rating for DIY: NVIDIA RTX PRO 6000 Blackwell Workstation Edition, 96GB + AMD EPYC 9015. (missing: measured prefill_tokens_sec, compute_tflops_fp16)	$42,159 (derived) DERIVEDHybrid · RTX PRO 6000 Blackwell Workstation Edition, 96GB + EPYC 9015 · 8× Server Premier KSM64R52BD4-96MC 96GB DDR5-6400 RDIMM observed priceArithmetic on observed inputs only. No tunable assumption enters it. observed_price = Σ (cheapest observed unit price × quantity) $598 + $839 + $3,627 × 8 + $10,795 + $70 + $180 + $660 = $42,159 AMD EPYC 9015 59844 ¢ — SabrePC · new · tier 2/sabrepc_ngstate ASRock Rack GENOAD8QM3-2T/BCM 83900 ¢ — Newegg · new · tier 3/newegg_realtime Kingston Server Premier KSM64R52BD4-96MC 96GB DDR5-6400 RDIMM 362697 ¢ — SabrePC · new · tier 2/sabrepc_ngstate NVIDIA RTX PRO 6000 Blackwell Workstation Edition, 96GB 1079547 ¢ — SabrePC · new · tier 2/sabrepc_ngstate ARCTIC Freezer 4U-SP5 6999 ¢ — ARCTIC · new · tier 1/arctic_jsonld Phanteks Enthoo Pro 2 Server Edition 17999 ¢ — Newegg · new · tier 3/newegg_realtime Seasonic PRIME PX-2200 65999 ¢ — Newegg · new · tier 3/newegg_realtime Listed price only: shipping and sales tax are NOT in this figure. One price per part — the cheapest monitored, non-quarantined observation.
Hybrid · RTX PRO 6000 Blackwell Max-Q Workstation Edition, 96GB + EPYC 9015 · 8× Server Premier KSM64R52BD4-96MC 96GB DDR5-6400 RDIMM 823 GB (derived) DERIVEDHybrid · RTX PRO 6000 Blackwell Max-Q Workstation Edition, 96GB + EPYC 9015 · 8× Server Premier KSM64R52BD4-96MC 96GB DDR5-6400 RDIMM usable memoryArithmetic on observed inputs only. No tunable assumption enters it. available_gb = Σ usable_gb over memory pools GPU VRAM 93.12 + System RAM 729.6 = 822.72 GB GPU VRAM.usable_gb 93.12 GB — 96 GB physical System RAM.usable_gb 729.5999999999999 GB — 768 GB physical Usable memory is below physical memory for at least one pool (OS/driver reservation; editable in Settings). usable 1,792 (observed) OBSERVEDHybrid · RTX PRO 6000 Blackwell Max-Q Workstation Edition, 96GB + EPYC 9015 · 8× Server Premier KSM64R52BD4-96MC 96GB DDR5-6400 RDIMM peak memory bandwidthMeasured or published, passed through unchanged — not computed here. peak memory bandwidth — published at the source, passed through unchanged; nothing is computed here 1,792 GB/s, as recorded from the GPU VRAM spec of Hybrid · RTX PRO 6000 Blackwell Max-Q Workstation Edition, 96GB + EPYC 9015 · 8× Server Premier KSM64R52BD4-96MC 96GB DDR5-6400 RDIMM peak_memory_bandwidth 1,792 GB/s — the GPU VRAM spec of Hybrid · RTX PRO 6000 Blackwell Max-Q Workstation Edition, 96GB + EPYC 9015 · 8× Server Premier KSM64R52BD4-96MC 96GB DDR5-6400 RDIMM Theoretical peak. Decode reaches a fraction of it — the efficiency factor is in the decode speed's own workings. The fastest pool. A machine with more than one pool places weights fastest-pool-first. Observed, not derived: this app records the figure and its source, and does not re-measure it. GB/s	Interactive	15.7 tok/s (estimate) ESTIMATEQwen3-VL 235B-A22B Instruct decode speed on Hybrid · RTX PRO 6000 Blackwell Max-Q Workstation Edition, 96GB + EPYC 9015 · 8× Server Premier KSM64R52BD4-96MC 96GB DDR5-6400 RDIMMArithmetic that also depends on a tunable assumption, listed below. tok_s = 1 / (Σ_pool bytes_per_token × share_pool / (bandwidth_pool × efficiency_pool) + Σ_pool share_pool × overhead_pool) GPU VRAM: 11.505 GB ÷ (1,792 GB/s × 0.48) + System RAM: 6.968 GB ÷ (307.2 GB/s × 0.7) + 17.81 ms fixed per-token overhead (share-weighted) → 15.73 tok/s GPU VRAM.bandwidth_gbs 1792 GB/s — component spec GPU VRAM.efficiency_moe 0.48 — Settings efficiency.gpu_vram — Discrete GPU: 10 dense + 3 MoE points; fit on published benchmarks, leave-one-out validated — docs/research/12b-estimator-validation.md GPU VRAM.active_share 0.6228 — placement System RAM.bandwidth_gbs 307.2 GB/s — component spec System RAM.efficiency_moe 0.7 — Settings efficiency.server_ddr5 — Server DDR5/MRDIMM single socket: 6 dense + 7 MoE points (small n); fit on published benchmarks, leave-one-out validated — docs/research/12b-estimator-validation.md. The 44.61 ms per-token overhead is a CPU-engine constant, not a property of DDR5: its fit is set by a 128-core Xeon 6980P running llama.cpp (55.3 ms/token on a 3B model) and SGLang (17.0 ms) with the CPU executing every layer (research 12 §6). A GPU spilling into this memory pays it only for the share of each token read from RAM (lib/perf/decode.ts refinement 3). System RAM.active_share 0.3772 — placement per_token_overhead_ms 17.811206136367083 ms — derived: 0.623 × 1.58 ms (GPU VRAM) + 0.377 × 44.61 ms (System RAM) = 17.81 ms bytes_per_token_gb 18.4738 GB MoE: only active parameters are read per token (SPEC §3). Effective bits include embeddings and higher-precision tensors in mixed quants. Assumes active weights are quantized like the file average. KV cache is re-read on every token (DECISIONS D-008). Pools are read sequentially per token, so per-pool times add (harmonic combination). Per-token overhead on a split placement is DERIVED, not measured: each pool's own single-pool figure is charged in proportion to the share of each token it serves, because the fixed cost is per operation and the split puts that share of the operations on each device. Calibrated with that pool serving the whole model (D-013): GPU VRAM (1.58 ms), System RAM (44.61 ms). Batch size 1 (single user, interactive).	Unavailable — No prefill measurement or compute rating for DIY: NVIDIA RTX PRO 6000 Blackwell Max-Q Workstation Edition, 96GB + AMD EPYC 9015. (missing: measured prefill_tokens_sec, compute_tflops_fp16)	$43,294 (derived) DERIVEDHybrid · RTX PRO 6000 Blackwell Max-Q Workstation Edition, 96GB + EPYC 9015 · 8× Server Premier KSM64R52BD4-96MC 96GB DDR5-6400 RDIMM observed priceArithmetic on observed inputs only. No tunable assumption enters it. observed_price = Σ (cheapest observed unit price × quantity) $598 + $839 + $3,627 × 8 + $11,931 + $70 + $180 + $660 = $43,294 AMD EPYC 9015 59844 ¢ — SabrePC · new · tier 2/sabrepc_ngstate ASRock Rack GENOAD8QM3-2T/BCM 83900 ¢ — Newegg · new · tier 3/newegg_realtime Kingston Server Premier KSM64R52BD4-96MC 96GB DDR5-6400 RDIMM 362697 ¢ — SabrePC · new · tier 2/sabrepc_ngstate NVIDIA RTX PRO 6000 Blackwell Max-Q Workstation Edition, 96GB 1193061 ¢ — SabrePC · new · tier 2/sabrepc_ngstate ARCTIC Freezer 4U-SP5 6999 ¢ — ARCTIC · new · tier 1/arctic_jsonld Phanteks Enthoo Pro 2 Server Edition 17999 ¢ — Newegg · new · tier 3/newegg_realtime Seasonic PRIME PX-2200 65999 ¢ — Newegg · new · tier 3/newegg_realtime Listed price only: shipping and sales tax are NOT in this figure. One price per part — the cheapest monitored, non-quarantined observation.
Hybrid · RTX PRO 6000 Blackwell Workstation Edition, 96GB + EPYC 9015 · 8× Server Premier KSM64R52BD4-128MB 128GB DDR5-6400 RDIMM 1,066 GB (derived) DERIVEDHybrid · RTX PRO 6000 Blackwell Workstation Edition, 96GB + EPYC 9015 · 8× Server Premier KSM64R52BD4-128MB 128GB DDR5-6400 RDIMM usable memoryArithmetic on observed inputs only. No tunable assumption enters it. available_gb = Σ usable_gb over memory pools GPU VRAM 93.12 + System RAM 972.8 = 1,065.92 GB GPU VRAM.usable_gb 93.12 GB — 96 GB physical System RAM.usable_gb 972.8 GB — 1,024 GB physical Usable memory is below physical memory for at least one pool (OS/driver reservation; editable in Settings). usable 1,792 (observed) OBSERVEDHybrid · RTX PRO 6000 Blackwell Workstation Edition, 96GB + EPYC 9015 · 8× Server Premier KSM64R52BD4-128MB 128GB DDR5-6400 RDIMM peak memory bandwidthMeasured or published, passed through unchanged — not computed here. peak memory bandwidth — published at the source, passed through unchanged; nothing is computed here 1,792 GB/s, as recorded from the GPU VRAM spec of Hybrid · RTX PRO 6000 Blackwell Workstation Edition, 96GB + EPYC 9015 · 8× Server Premier KSM64R52BD4-128MB 128GB DDR5-6400 RDIMM peak_memory_bandwidth 1,792 GB/s — the GPU VRAM spec of Hybrid · RTX PRO 6000 Blackwell Workstation Edition, 96GB + EPYC 9015 · 8× Server Premier KSM64R52BD4-128MB 128GB DDR5-6400 RDIMM Theoretical peak. Decode reaches a fraction of it — the efficiency factor is in the decode speed's own workings. The fastest pool. A machine with more than one pool places weights fastest-pool-first. Observed, not derived: this app records the figure and its source, and does not re-measure it. GB/s	Interactive	15.7 tok/s (estimate) ESTIMATEQwen3-VL 235B-A22B Instruct decode speed on Hybrid · RTX PRO 6000 Blackwell Workstation Edition, 96GB + EPYC 9015 · 8× Server Premier KSM64R52BD4-128MB 128GB DDR5-6400 RDIMMArithmetic that also depends on a tunable assumption, listed below. tok_s = 1 / (Σ_pool bytes_per_token × share_pool / (bandwidth_pool × efficiency_pool) + Σ_pool share_pool × overhead_pool) GPU VRAM: 11.505 GB ÷ (1,792 GB/s × 0.48) + System RAM: 6.968 GB ÷ (307.2 GB/s × 0.7) + 17.81 ms fixed per-token overhead (share-weighted) → 15.73 tok/s GPU VRAM.bandwidth_gbs 1792 GB/s — component spec GPU VRAM.efficiency_moe 0.48 — Settings efficiency.gpu_vram — Discrete GPU: 10 dense + 3 MoE points; fit on published benchmarks, leave-one-out validated — docs/research/12b-estimator-validation.md GPU VRAM.active_share 0.6228 — placement System RAM.bandwidth_gbs 307.2 GB/s — component spec System RAM.efficiency_moe 0.7 — Settings efficiency.server_ddr5 — Server DDR5/MRDIMM single socket: 6 dense + 7 MoE points (small n); fit on published benchmarks, leave-one-out validated — docs/research/12b-estimator-validation.md. The 44.61 ms per-token overhead is a CPU-engine constant, not a property of DDR5: its fit is set by a 128-core Xeon 6980P running llama.cpp (55.3 ms/token on a 3B model) and SGLang (17.0 ms) with the CPU executing every layer (research 12 §6). A GPU spilling into this memory pays it only for the share of each token read from RAM (lib/perf/decode.ts refinement 3). System RAM.active_share 0.3772 — placement per_token_overhead_ms 17.811206136367083 ms — derived: 0.623 × 1.58 ms (GPU VRAM) + 0.377 × 44.61 ms (System RAM) = 17.81 ms bytes_per_token_gb 18.4738 GB MoE: only active parameters are read per token (SPEC §3). Effective bits include embeddings and higher-precision tensors in mixed quants. Assumes active weights are quantized like the file average. KV cache is re-read on every token (DECISIONS D-008). Pools are read sequentially per token, so per-pool times add (harmonic combination). Per-token overhead on a split placement is DERIVED, not measured: each pool's own single-pool figure is charged in proportion to the share of each token it serves, because the fixed cost is per operation and the split puts that share of the operations on each device. Calibrated with that pool serving the whole model (D-013): GPU VRAM (1.58 ms), System RAM (44.61 ms). Batch size 1 (single user, interactive).	Unavailable — No prefill measurement or compute rating for DIY: NVIDIA RTX PRO 6000 Blackwell Workstation Edition, 96GB + AMD EPYC 9015. (missing: measured prefill_tokens_sec, compute_tflops_fp16)	$44,568 (derived) DERIVEDHybrid · RTX PRO 6000 Blackwell Workstation Edition, 96GB + EPYC 9015 · 8× Server Premier KSM64R52BD4-128MB 128GB DDR5-6400 RDIMM observed priceArithmetic on observed inputs only. No tunable assumption enters it. observed_price = Σ (cheapest observed unit price × quantity) $598 + $839 + $3,928 × 8 + $10,795 + $70 + $180 + $660 = $44,568 AMD EPYC 9015 59844 ¢ — SabrePC · new · tier 2/sabrepc_ngstate ASRock Rack GENOAD8QM3-2T/BCM 83900 ¢ — Newegg · new · tier 3/newegg_realtime Kingston Server Premier KSM64R52BD4-128MB 128GB DDR5-6400 RDIMM 392811 ¢ — SabrePC · new · tier 2/sabrepc_ngstate NVIDIA RTX PRO 6000 Blackwell Workstation Edition, 96GB 1079547 ¢ — SabrePC · new · tier 2/sabrepc_ngstate ARCTIC Freezer 4U-SP5 6999 ¢ — ARCTIC · new · tier 1/arctic_jsonld Phanteks Enthoo Pro 2 Server Edition 17999 ¢ — Newegg · new · tier 3/newegg_realtime Seasonic PRIME PX-2200 65999 ¢ — Newegg · new · tier 3/newegg_realtime Listed price only: shipping and sales tax are NOT in this figure. One price per part — the cheapest monitored, non-quarantined observation.
Hybrid · RTX PRO 6000 Blackwell Max-Q Workstation Edition, 96GB + EPYC 9015 · 8× Server Premier KSM64R52BD4-128MB 128GB DDR5-6400 RDIMM 1,066 GB (derived) DERIVEDHybrid · RTX PRO 6000 Blackwell Max-Q Workstation Edition, 96GB + EPYC 9015 · 8× Server Premier KSM64R52BD4-128MB 128GB DDR5-6400 RDIMM usable memoryArithmetic on observed inputs only. No tunable assumption enters it. available_gb = Σ usable_gb over memory pools GPU VRAM 93.12 + System RAM 972.8 = 1,065.92 GB GPU VRAM.usable_gb 93.12 GB — 96 GB physical System RAM.usable_gb 972.8 GB — 1,024 GB physical Usable memory is below physical memory for at least one pool (OS/driver reservation; editable in Settings). usable 1,792 (observed) OBSERVEDHybrid · RTX PRO 6000 Blackwell Max-Q Workstation Edition, 96GB + EPYC 9015 · 8× Server Premier KSM64R52BD4-128MB 128GB DDR5-6400 RDIMM peak memory bandwidthMeasured or published, passed through unchanged — not computed here. peak memory bandwidth — published at the source, passed through unchanged; nothing is computed here 1,792 GB/s, as recorded from the GPU VRAM spec of Hybrid · RTX PRO 6000 Blackwell Max-Q Workstation Edition, 96GB + EPYC 9015 · 8× Server Premier KSM64R52BD4-128MB 128GB DDR5-6400 RDIMM peak_memory_bandwidth 1,792 GB/s — the GPU VRAM spec of Hybrid · RTX PRO 6000 Blackwell Max-Q Workstation Edition, 96GB + EPYC 9015 · 8× Server Premier KSM64R52BD4-128MB 128GB DDR5-6400 RDIMM Theoretical peak. Decode reaches a fraction of it — the efficiency factor is in the decode speed's own workings. The fastest pool. A machine with more than one pool places weights fastest-pool-first. Observed, not derived: this app records the figure and its source, and does not re-measure it. GB/s	Interactive	15.7 tok/s (estimate) ESTIMATEQwen3-VL 235B-A22B Instruct decode speed on Hybrid · RTX PRO 6000 Blackwell Max-Q Workstation Edition, 96GB + EPYC 9015 · 8× Server Premier KSM64R52BD4-128MB 128GB DDR5-6400 RDIMMArithmetic that also depends on a tunable assumption, listed below. tok_s = 1 / (Σ_pool bytes_per_token × share_pool / (bandwidth_pool × efficiency_pool) + Σ_pool share_pool × overhead_pool) GPU VRAM: 11.505 GB ÷ (1,792 GB/s × 0.48) + System RAM: 6.968 GB ÷ (307.2 GB/s × 0.7) + 17.81 ms fixed per-token overhead (share-weighted) → 15.73 tok/s GPU VRAM.bandwidth_gbs 1792 GB/s — component spec GPU VRAM.efficiency_moe 0.48 — Settings efficiency.gpu_vram — Discrete GPU: 10 dense + 3 MoE points; fit on published benchmarks, leave-one-out validated — docs/research/12b-estimator-validation.md GPU VRAM.active_share 0.6228 — placement System RAM.bandwidth_gbs 307.2 GB/s — component spec System RAM.efficiency_moe 0.7 — Settings efficiency.server_ddr5 — Server DDR5/MRDIMM single socket: 6 dense + 7 MoE points (small n); fit on published benchmarks, leave-one-out validated — docs/research/12b-estimator-validation.md. The 44.61 ms per-token overhead is a CPU-engine constant, not a property of DDR5: its fit is set by a 128-core Xeon 6980P running llama.cpp (55.3 ms/token on a 3B model) and SGLang (17.0 ms) with the CPU executing every layer (research 12 §6). A GPU spilling into this memory pays it only for the share of each token read from RAM (lib/perf/decode.ts refinement 3). System RAM.active_share 0.3772 — placement per_token_overhead_ms 17.811206136367083 ms — derived: 0.623 × 1.58 ms (GPU VRAM) + 0.377 × 44.61 ms (System RAM) = 17.81 ms bytes_per_token_gb 18.4738 GB MoE: only active parameters are read per token (SPEC §3). Effective bits include embeddings and higher-precision tensors in mixed quants. Assumes active weights are quantized like the file average. KV cache is re-read on every token (DECISIONS D-008). Pools are read sequentially per token, so per-pool times add (harmonic combination). Per-token overhead on a split placement is DERIVED, not measured: each pool's own single-pool figure is charged in proportion to the share of each token it serves, because the fixed cost is per operation and the split puts that share of the operations on each device. Calibrated with that pool serving the whole model (D-013): GPU VRAM (1.58 ms), System RAM (44.61 ms). Batch size 1 (single user, interactive).	Unavailable — No prefill measurement or compute rating for DIY: NVIDIA RTX PRO 6000 Blackwell Max-Q Workstation Edition, 96GB + AMD EPYC 9015. (missing: measured prefill_tokens_sec, compute_tflops_fp16)	$45,703 (derived) DERIVEDHybrid · RTX PRO 6000 Blackwell Max-Q Workstation Edition, 96GB + EPYC 9015 · 8× Server Premier KSM64R52BD4-128MB 128GB DDR5-6400 RDIMM observed priceArithmetic on observed inputs only. No tunable assumption enters it. observed_price = Σ (cheapest observed unit price × quantity) $598 + $839 + $3,928 × 8 + $11,931 + $70 + $180 + $660 = $45,703 AMD EPYC 9015 59844 ¢ — SabrePC · new · tier 2/sabrepc_ngstate ASRock Rack GENOAD8QM3-2T/BCM 83900 ¢ — Newegg · new · tier 3/newegg_realtime Kingston Server Premier KSM64R52BD4-128MB 128GB DDR5-6400 RDIMM 392811 ¢ — SabrePC · new · tier 2/sabrepc_ngstate NVIDIA RTX PRO 6000 Blackwell Max-Q Workstation Edition, 96GB 1193061 ¢ — SabrePC · new · tier 2/sabrepc_ngstate ARCTIC Freezer 4U-SP5 6999 ¢ — ARCTIC · new · tier 1/arctic_jsonld Phanteks Enthoo Pro 2 Server Edition 17999 ¢ — Newegg · new · tier 3/newegg_realtime Seasonic PRIME PX-2200 65999 ¢ — Newegg · new · tier 3/newegg_realtime Listed price only: shipping and sales tax are NOT in this figure. One price per part — the cheapest monitored, non-quarantined observation.
2× GeForce RTX 5090 32G GAMING TRIO OC · EPYC 9015 · 8× M321R4GA3BB6-CQK 32GB DDR5-4800 RDIMM 305 GB (derived) DERIVED2× GeForce RTX 5090 32G GAMING TRIO OC · EPYC 9015 · 8× M321R4GA3BB6-CQK 32GB DDR5-4800 RDIMM usable memoryArithmetic on observed inputs only. No tunable assumption enters it. available_gb = Σ usable_gb over memory pools GPU VRAM 62.08 + System RAM 243.2 = 305.28 GB GPU VRAM.usable_gb 62.08 GB — 64 GB physical System RAM.usable_gb 243.2 GB — 256 GB physical Usable memory is below physical memory for at least one pool (OS/driver reservation; editable in Settings). usable 1,792 (observed) OBSERVED2× GeForce RTX 5090 32G GAMING TRIO OC · EPYC 9015 · 8× M321R4GA3BB6-CQK 32GB DDR5-4800 RDIMM peak memory bandwidthMeasured or published, passed through unchanged — not computed here. peak memory bandwidth — published at the source, passed through unchanged; nothing is computed here 1,792 GB/s, as recorded from the GPU VRAM spec of 2× GeForce RTX 5090 32G GAMING TRIO OC · EPYC 9015 · 8× M321R4GA3BB6-CQK 32GB DDR5-4800 RDIMM peak_memory_bandwidth 1,792 GB/s — the GPU VRAM spec of 2× GeForce RTX 5090 32G GAMING TRIO OC · EPYC 9015 · 8× M321R4GA3BB6-CQK 32GB DDR5-4800 RDIMM Theoretical peak. Decode reaches a fraction of it — the efficiency factor is in the decode speed's own workings. The fastest pool. A machine with more than one pool places weights fastest-pool-first. Observed, not derived: this app records the figure and its source, and does not re-measure it. GB/s	Usable	11.6 tok/s (estimate) ESTIMATEQwen3-VL 235B-A22B Instruct decode speed on 2× GeForce RTX 5090 32G GAMING TRIO OC · EPYC 9015 · 8× M321R4GA3BB6-CQK 32GB DDR5-4800 RDIMMArithmetic that also depends on a tunable assumption, listed below. tok_s = 1 / (Σ_pool bytes_per_token × share_pool / (bandwidth_pool × efficiency_pool) + Σ_pool share_pool × overhead_pool) GPU VRAM: 7.67 GB ÷ (1,792 GB/s × 0.48) + System RAM: 10.804 GB ÷ (307.2 GB/s × 0.7) + 26.74 ms fixed per-token overhead (share-weighted) → 11.64 tok/s GPU VRAM.bandwidth_gbs 1792 GB/s — component spec GPU VRAM.efficiency_moe 0.48 — Settings efficiency.gpu_vram — Discrete GPU: 10 dense + 3 MoE points; fit on published benchmarks, leave-one-out validated — docs/research/12b-estimator-validation.md GPU VRAM.active_share 0.4152 — placement System RAM.bandwidth_gbs 307.2 GB/s — component spec System RAM.efficiency_moe 0.7 — Settings efficiency.server_ddr5 — Server DDR5/MRDIMM single socket: 6 dense + 7 MoE points (small n); fit on published benchmarks, leave-one-out validated — docs/research/12b-estimator-validation.md. The 44.61 ms per-token overhead is a CPU-engine constant, not a property of DDR5: its fit is set by a 128-core Xeon 6980P running llama.cpp (55.3 ms/token on a 3B model) and SGLang (17.0 ms) with the CPU executing every layer (research 12 §6). A GPU spilling into this memory pays it only for the share of each token read from RAM (lib/perf/decode.ts refinement 3). System RAM.active_share 0.5848 — placement per_token_overhead_ms 26.744137424244723 ms — derived: 0.415 × 1.58 ms (GPU VRAM) + 0.585 × 44.61 ms (System RAM) = 26.74 ms bytes_per_token_gb 18.4738 GB MoE: only active parameters are read per token (SPEC §3). Effective bits include embeddings and higher-precision tensors in mixed quants. Assumes active weights are quantized like the file average. KV cache is re-read on every token (DECISIONS D-008). Pools are read sequentially per token, so per-pool times add (harmonic combination). layer split across 2 GPUs: capacity adds, per-token bandwidth does not (layers run sequentially) Per-token overhead on a split placement is DERIVED, not measured: each pool's own single-pool figure is charged in proportion to the share of each token it serves, because the fixed cost is per operation and the split puts that share of the operations on each device. Calibrated with that pool serving the whole model (D-013): GPU VRAM (1.58 ms), System RAM (44.61 ms). Batch size 1 (single user, interactive).	19.3 s (estimate) ESTIMATEQwen3-VL 235B-A22B Instruct time to first token on 2× GeForce RTX 5090 32G GAMING TRIO OC · EPYC 9015 · 8× M321R4GA3BB6-CQK 32GB DDR5-4800 RDIMMArithmetic that also depends on a tunable assumption, listed below. ttft_s = prompt_tokens / prefill_tok_s = 32,768 / 1,699.82 = 19.3 s prompt_tokens 32768 tokens prefill_tok_s 1699.82 tok/s No published prefill measurement for this hardware; scaled from peak compute (estimate). Model q_dim unknown — attention FLOPs omitted, so long-prompt TTFT is understated.	$14,262 (derived) DERIVED2× GeForce RTX 5090 32G GAMING TRIO OC · EPYC 9015 · 8× M321R4GA3BB6-CQK 32GB DDR5-4800 RDIMM observed priceArithmetic on observed inputs only. No tunable assumption enters it. observed_price = Σ (cheapest observed unit price × quantity) $598 + $839 + $214 × 8 + $5,100 × 2 + $70 + $180 + $660 = $14,262 AMD EPYC 9015 59844 ¢ — SabrePC · new · tier 2/sabrepc_ngstate ASRock Rack GENOAD8QM3-2T/BCM 83900 ¢ — Newegg · new · tier 3/newegg_realtime Samsung M321R4GA3BB6-CQK 32GB DDR5-4800 RDIMM 21427 ¢ — SabrePC · new · tier 2/sabrepc_ngstate MSI GeForce RTX 5090 32G GAMING TRIO OC 509999 ¢ — Newegg · new · tier 3/newegg_realtime ARCTIC Freezer 4U-SP5 6999 ¢ — ARCTIC · new · tier 1/arctic_jsonld Phanteks Enthoo Pro 2 Server Edition 17999 ¢ — Newegg · new · tier 3/newegg_realtime Seasonic PRIME PX-2200 65999 ¢ — Newegg · new · tier 3/newegg_realtime Listed price only: shipping and sales tax are NOT in this figure. One price per part — the cheapest monitored, non-quarantined observation.
3× Radeon AI PRO R9700 Creator 32GB (R9700 CT) · EPYC 9015 · 8× M321R4GA3BB6-CQK 32GB DDR5-4800 RDIMM 336 GB (derived) DERIVED3× Radeon AI PRO R9700 Creator 32GB (R9700 CT) · EPYC 9015 · 8× M321R4GA3BB6-CQK 32GB DDR5-4800 RDIMM usable memoryArithmetic on observed inputs only. No tunable assumption enters it. available_gb = Σ usable_gb over memory pools GPU VRAM 93.12 + System RAM 243.2 = 336.32 GB GPU VRAM.usable_gb 93.12 GB — 96 GB physical System RAM.usable_gb 243.2 GB — 256 GB physical Usable memory is below physical memory for at least one pool (OS/driver reservation; editable in Settings). usable 640 (observed) OBSERVED3× Radeon AI PRO R9700 Creator 32GB (R9700 CT) · EPYC 9015 · 8× M321R4GA3BB6-CQK 32GB DDR5-4800 RDIMM peak memory bandwidthMeasured or published, passed through unchanged — not computed here. peak memory bandwidth — published at the source, passed through unchanged; nothing is computed here 640 GB/s, as recorded from the GPU VRAM spec of 3× Radeon AI PRO R9700 Creator 32GB (R9700 CT) · EPYC 9015 · 8× M321R4GA3BB6-CQK 32GB DDR5-4800 RDIMM peak_memory_bandwidth 640 GB/s — the GPU VRAM spec of 3× Radeon AI PRO R9700 Creator 32GB (R9700 CT) · EPYC 9015 · 8× M321R4GA3BB6-CQK 32GB DDR5-4800 RDIMM Theoretical peak. Decode reaches a fraction of it — the efficiency factor is in the decode speed's own workings. The fastest pool. A machine with more than one pool places weights fastest-pool-first. Observed, not derived: this app records the figure and its source, and does not re-measure it. GB/s	Usable	11.4 tok/s (estimate) ESTIMATEQwen3-VL 235B-A22B Instruct decode speed on 3× Radeon AI PRO R9700 Creator 32GB (R9700 CT) · EPYC 9015 · 8× M321R4GA3BB6-CQK 32GB DDR5-4800 RDIMMArithmetic that also depends on a tunable assumption, listed below. tok_s = 1 / (Σ_pool bytes_per_token × share_pool / (bandwidth_pool × efficiency_pool) + Σ_pool share_pool × overhead_pool) GPU VRAM: 11.505 GB ÷ (640 GB/s × 0.48) + System RAM: 6.968 GB ÷ (307.2 GB/s × 0.7) + 17.81 ms fixed per-token overhead (share-weighted) → 11.41 tok/s GPU VRAM.bandwidth_gbs 640 GB/s — component spec GPU VRAM.efficiency_moe 0.48 — Settings efficiency.gpu_vram — Discrete GPU: 10 dense + 3 MoE points; fit on published benchmarks, leave-one-out validated — docs/research/12b-estimator-validation.md GPU VRAM.active_share 0.6228 — placement System RAM.bandwidth_gbs 307.2 GB/s — component spec System RAM.efficiency_moe 0.7 — Settings efficiency.server_ddr5 — Server DDR5/MRDIMM single socket: 6 dense + 7 MoE points (small n); fit on published benchmarks, leave-one-out validated — docs/research/12b-estimator-validation.md. The 44.61 ms per-token overhead is a CPU-engine constant, not a property of DDR5: its fit is set by a 128-core Xeon 6980P running llama.cpp (55.3 ms/token on a 3B model) and SGLang (17.0 ms) with the CPU executing every layer (research 12 §6). A GPU spilling into this memory pays it only for the share of each token read from RAM (lib/perf/decode.ts refinement 3). System RAM.active_share 0.3772 — placement per_token_overhead_ms 17.811206136367083 ms — derived: 0.623 × 1.58 ms (GPU VRAM) + 0.377 × 44.61 ms (System RAM) = 17.81 ms bytes_per_token_gb 18.4738 GB MoE: only active parameters are read per token (SPEC §3). Effective bits include embeddings and higher-precision tensors in mixed quants. Assumes active weights are quantized like the file average. KV cache is re-read on every token (DECISIONS D-008). Pools are read sequentially per token, so per-pool times add (harmonic combination). layer split across 3 GPUs: capacity adds, per-token bandwidth does not (layers run sequentially) Per-token overhead on a split placement is DERIVED, not measured: each pool's own single-pool figure is charged in proportion to the share of each token it serves, because the fixed cost is per operation and the split puts that share of the operations on each device. Calibrated with that pool serving the whole model (D-013): GPU VRAM (1.58 ms), System RAM (44.61 ms). Batch size 1 (single user, interactive).	21.1 s (estimate) ESTIMATEQwen3-VL 235B-A22B Instruct time to first token on 3× Radeon AI PRO R9700 Creator 32GB (R9700 CT) · EPYC 9015 · 8× M321R4GA3BB6-CQK 32GB DDR5-4800 RDIMMArithmetic that also depends on a tunable assumption, listed below. ttft_s = prompt_tokens / prefill_tok_s = 32,768 / 1,549.72 = 21.1 s prompt_tokens 32768 tokens prefill_tok_s 1549.72 tok/s No published prefill measurement for this hardware; scaled from peak compute (estimate). Model q_dim unknown — attention FLOPs omitted, so long-prompt TTFT is understated.	$9,162 (derived) DERIVED3× Radeon AI PRO R9700 Creator 32GB (R9700 CT) · EPYC 9015 · 8× M321R4GA3BB6-CQK 32GB DDR5-4800 RDIMM observed priceArithmetic on observed inputs only. No tunable assumption enters it. observed_price = Σ (cheapest observed unit price × quantity) $598 + $839 + $214 × 8 + $1,700 × 3 + $70 + $180 + $660 = $9,162 AMD EPYC 9015 59844 ¢ — SabrePC · new · tier 2/sabrepc_ngstate ASRock Rack GENOAD8QM3-2T/BCM 83900 ¢ — Newegg · new · tier 3/newegg_realtime Samsung M321R4GA3BB6-CQK 32GB DDR5-4800 RDIMM 21427 ¢ — SabrePC · new · tier 2/sabrepc_ngstate ASRock Radeon AI PRO R9700 Creator 32GB (R9700 CT) 169999 ¢ — Newegg · new · tier 2/seed:research ARCTIC Freezer 4U-SP5 6999 ¢ — ARCTIC · new · tier 1/arctic_jsonld Phanteks Enthoo Pro 2 Server Edition 17999 ¢ — Newegg · new · tier 3/newegg_realtime Seasonic PRIME PX-2200 65999 ¢ — Newegg · new · tier 3/newegg_realtime Listed price only: shipping and sales tax are NOT in this figure. One price per part — the cheapest monitored, non-quarantined observation.
3× Radeon AI PRO R9700 32GB (32358-01-20G) · EPYC 9015 · 8× M321R4GA3BB6-CQK 32GB DDR5-4800 RDIMM 336 GB (derived) DERIVED3× Radeon AI PRO R9700 32GB (32358-01-20G) · EPYC 9015 · 8× M321R4GA3BB6-CQK 32GB DDR5-4800 RDIMM usable memoryArithmetic on observed inputs only. No tunable assumption enters it. available_gb = Σ usable_gb over memory pools GPU VRAM 93.12 + System RAM 243.2 = 336.32 GB GPU VRAM.usable_gb 93.12 GB — 96 GB physical System RAM.usable_gb 243.2 GB — 256 GB physical Usable memory is below physical memory for at least one pool (OS/driver reservation; editable in Settings). usable 640 (observed) OBSERVED3× Radeon AI PRO R9700 32GB (32358-01-20G) · EPYC 9015 · 8× M321R4GA3BB6-CQK 32GB DDR5-4800 RDIMM peak memory bandwidthMeasured or published, passed through unchanged — not computed here. peak memory bandwidth — published at the source, passed through unchanged; nothing is computed here 640 GB/s, as recorded from the GPU VRAM spec of 3× Radeon AI PRO R9700 32GB (32358-01-20G) · EPYC 9015 · 8× M321R4GA3BB6-CQK 32GB DDR5-4800 RDIMM peak_memory_bandwidth 640 GB/s — the GPU VRAM spec of 3× Radeon AI PRO R9700 32GB (32358-01-20G) · EPYC 9015 · 8× M321R4GA3BB6-CQK 32GB DDR5-4800 RDIMM Theoretical peak. Decode reaches a fraction of it — the efficiency factor is in the decode speed's own workings. The fastest pool. A machine with more than one pool places weights fastest-pool-first. Observed, not derived: this app records the figure and its source, and does not re-measure it. GB/s	Usable	11.4 tok/s (estimate) ESTIMATEQwen3-VL 235B-A22B Instruct decode speed on 3× Radeon AI PRO R9700 32GB (32358-01-20G) · EPYC 9015 · 8× M321R4GA3BB6-CQK 32GB DDR5-4800 RDIMMArithmetic that also depends on a tunable assumption, listed below. tok_s = 1 / (Σ_pool bytes_per_token × share_pool / (bandwidth_pool × efficiency_pool) + Σ_pool share_pool × overhead_pool) GPU VRAM: 11.505 GB ÷ (640 GB/s × 0.48) + System RAM: 6.968 GB ÷ (307.2 GB/s × 0.7) + 17.81 ms fixed per-token overhead (share-weighted) → 11.41 tok/s GPU VRAM.bandwidth_gbs 640 GB/s — component spec GPU VRAM.efficiency_moe 0.48 — Settings efficiency.gpu_vram — Discrete GPU: 10 dense + 3 MoE points; fit on published benchmarks, leave-one-out validated — docs/research/12b-estimator-validation.md GPU VRAM.active_share 0.6228 — placement System RAM.bandwidth_gbs 307.2 GB/s — component spec System RAM.efficiency_moe 0.7 — Settings efficiency.server_ddr5 — Server DDR5/MRDIMM single socket: 6 dense + 7 MoE points (small n); fit on published benchmarks, leave-one-out validated — docs/research/12b-estimator-validation.md. The 44.61 ms per-token overhead is a CPU-engine constant, not a property of DDR5: its fit is set by a 128-core Xeon 6980P running llama.cpp (55.3 ms/token on a 3B model) and SGLang (17.0 ms) with the CPU executing every layer (research 12 §6). A GPU spilling into this memory pays it only for the share of each token read from RAM (lib/perf/decode.ts refinement 3). System RAM.active_share 0.3772 — placement per_token_overhead_ms 17.811206136367083 ms — derived: 0.623 × 1.58 ms (GPU VRAM) + 0.377 × 44.61 ms (System RAM) = 17.81 ms bytes_per_token_gb 18.4738 GB MoE: only active parameters are read per token (SPEC §3). Effective bits include embeddings and higher-precision tensors in mixed quants. Assumes active weights are quantized like the file average. KV cache is re-read on every token (DECISIONS D-008). Pools are read sequentially per token, so per-pool times add (harmonic combination). layer split across 3 GPUs: capacity adds, per-token bandwidth does not (layers run sequentially) Per-token overhead on a split placement is DERIVED, not measured: each pool's own single-pool figure is charged in proportion to the share of each token it serves, because the fixed cost is per operation and the split puts that share of the operations on each device. Calibrated with that pool serving the whole model (D-013): GPU VRAM (1.58 ms), System RAM (44.61 ms). Batch size 1 (single user, interactive).	21.1 s (estimate) ESTIMATEQwen3-VL 235B-A22B Instruct time to first token on 3× Radeon AI PRO R9700 32GB (32358-01-20G) · EPYC 9015 · 8× M321R4GA3BB6-CQK 32GB DDR5-4800 RDIMMArithmetic that also depends on a tunable assumption, listed below. ttft_s = prompt_tokens / prefill_tok_s = 32,768 / 1,549.72 = 21.1 s prompt_tokens 32768 tokens prefill_tok_s 1549.72 tok/s No published prefill measurement for this hardware; scaled from peak compute (estimate). Model q_dim unknown — attention FLOPs omitted, so long-prompt TTFT is understated.	$8,875 (derived) DERIVED3× Radeon AI PRO R9700 32GB (32358-01-20G) · EPYC 9015 · 8× M321R4GA3BB6-CQK 32GB DDR5-4800 RDIMM observed priceArithmetic on observed inputs only. No tunable assumption enters it. observed_price = Σ (cheapest observed unit price × quantity) $598 + $839 + $214 × 8 + $1,604 × 3 + $70 + $180 + $660 = $8,875 AMD EPYC 9015 59844 ¢ — SabrePC · new · tier 2/sabrepc_ngstate ASRock Rack GENOAD8QM3-2T/BCM 83900 ¢ — Newegg · new · tier 3/newegg_realtime Samsung M321R4GA3BB6-CQK 32GB DDR5-4800 RDIMM 21427 ¢ — SabrePC · new · tier 2/sabrepc_ngstate Sapphire Radeon AI PRO R9700 32GB (32358-01-20G) 160431 ¢ — SabrePC · new · tier 2/sabrepc_ngstate ARCTIC Freezer 4U-SP5 6999 ¢ — ARCTIC · new · tier 1/arctic_jsonld Phanteks Enthoo Pro 2 Server Edition 17999 ¢ — Newegg · new · tier 3/newegg_realtime Seasonic PRIME PX-2200 65999 ¢ — Newegg · new · tier 3/newegg_realtime Listed price only: shipping and sales tax are NOT in this figure. One price per part — the cheapest monitored, non-quarantined observation.
3× AI TOP Radeon AI PRO R9700 32GB (GV-R9700AI TOP-32GD) · EPYC 9015 · 8× M321R4GA3BB6-CQK 32GB DDR5-4800 RDIMM 336 GB (derived) DERIVED3× AI TOP Radeon AI PRO R9700 32GB (GV-R9700AI TOP-32GD) · EPYC 9015 · 8× M321R4GA3BB6-CQK 32GB DDR5-4800 RDIMM usable memoryArithmetic on observed inputs only. No tunable assumption enters it. available_gb = Σ usable_gb over memory pools GPU VRAM 93.12 + System RAM 243.2 = 336.32 GB GPU VRAM.usable_gb 93.12 GB — 96 GB physical System RAM.usable_gb 243.2 GB — 256 GB physical Usable memory is below physical memory for at least one pool (OS/driver reservation; editable in Settings). usable 640 (observed) OBSERVED3× AI TOP Radeon AI PRO R9700 32GB (GV-R9700AI TOP-32GD) · EPYC 9015 · 8× M321R4GA3BB6-CQK 32GB DDR5-4800 RDIMM peak memory bandwidthMeasured or published, passed through unchanged — not computed here. peak memory bandwidth — published at the source, passed through unchanged; nothing is computed here 640 GB/s, as recorded from the GPU VRAM spec of 3× AI TOP Radeon AI PRO R9700 32GB (GV-R9700AI TOP-32GD) · EPYC 9015 · 8× M321R4GA3BB6-CQK 32GB DDR5-4800 RDIMM peak_memory_bandwidth 640 GB/s — the GPU VRAM spec of 3× AI TOP Radeon AI PRO R9700 32GB (GV-R9700AI TOP-32GD) · EPYC 9015 · 8× M321R4GA3BB6-CQK 32GB DDR5-4800 RDIMM Theoretical peak. Decode reaches a fraction of it — the efficiency factor is in the decode speed's own workings. The fastest pool. A machine with more than one pool places weights fastest-pool-first. Observed, not derived: this app records the figure and its source, and does not re-measure it. GB/s	Usable	11.4 tok/s (estimate) ESTIMATEQwen3-VL 235B-A22B Instruct decode speed on 3× AI TOP Radeon AI PRO R9700 32GB (GV-R9700AI TOP-32GD) · EPYC 9015 · 8× M321R4GA3BB6-CQK 32GB DDR5-4800 RDIMMArithmetic that also depends on a tunable assumption, listed below. tok_s = 1 / (Σ_pool bytes_per_token × share_pool / (bandwidth_pool × efficiency_pool) + Σ_pool share_pool × overhead_pool) GPU VRAM: 11.505 GB ÷ (640 GB/s × 0.48) + System RAM: 6.968 GB ÷ (307.2 GB/s × 0.7) + 17.81 ms fixed per-token overhead (share-weighted) → 11.41 tok/s GPU VRAM.bandwidth_gbs 640 GB/s — component spec GPU VRAM.efficiency_moe 0.48 — Settings efficiency.gpu_vram — Discrete GPU: 10 dense + 3 MoE points; fit on published benchmarks, leave-one-out validated — docs/research/12b-estimator-validation.md GPU VRAM.active_share 0.6228 — placement System RAM.bandwidth_gbs 307.2 GB/s — component spec System RAM.efficiency_moe 0.7 — Settings efficiency.server_ddr5 — Server DDR5/MRDIMM single socket: 6 dense + 7 MoE points (small n); fit on published benchmarks, leave-one-out validated — docs/research/12b-estimator-validation.md. The 44.61 ms per-token overhead is a CPU-engine constant, not a property of DDR5: its fit is set by a 128-core Xeon 6980P running llama.cpp (55.3 ms/token on a 3B model) and SGLang (17.0 ms) with the CPU executing every layer (research 12 §6). A GPU spilling into this memory pays it only for the share of each token read from RAM (lib/perf/decode.ts refinement 3). System RAM.active_share 0.3772 — placement per_token_overhead_ms 17.811206136367083 ms — derived: 0.623 × 1.58 ms (GPU VRAM) + 0.377 × 44.61 ms (System RAM) = 17.81 ms bytes_per_token_gb 18.4738 GB MoE: only active parameters are read per token (SPEC §3). Effective bits include embeddings and higher-precision tensors in mixed quants. Assumes active weights are quantized like the file average. KV cache is re-read on every token (DECISIONS D-008). Pools are read sequentially per token, so per-pool times add (harmonic combination). layer split across 3 GPUs: capacity adds, per-token bandwidth does not (layers run sequentially) Per-token overhead on a split placement is DERIVED, not measured: each pool's own single-pool figure is charged in proportion to the share of each token it serves, because the fixed cost is per operation and the split puts that share of the operations on each device. Calibrated with that pool serving the whole model (D-013): GPU VRAM (1.58 ms), System RAM (44.61 ms). Batch size 1 (single user, interactive).	21.1 s (estimate) ESTIMATEQwen3-VL 235B-A22B Instruct time to first token on 3× AI TOP Radeon AI PRO R9700 32GB (GV-R9700AI TOP-32GD) · EPYC 9015 · 8× M321R4GA3BB6-CQK 32GB DDR5-4800 RDIMMArithmetic that also depends on a tunable assumption, listed below. ttft_s = prompt_tokens / prefill_tok_s = 32,768 / 1,549.72 = 21.1 s prompt_tokens 32768 tokens prefill_tok_s 1549.72 tok/s No published prefill measurement for this hardware; scaled from peak compute (estimate). Model q_dim unknown — attention FLOPs omitted, so long-prompt TTFT is understated.	$9,194 (derived) DERIVED3× AI TOP Radeon AI PRO R9700 32GB (GV-R9700AI TOP-32GD) · EPYC 9015 · 8× M321R4GA3BB6-CQK 32GB DDR5-4800 RDIMM observed priceArithmetic on observed inputs only. No tunable assumption enters it. observed_price = Σ (cheapest observed unit price × quantity) $598 + $839 + $214 × 8 + $1,711 × 3 + $70 + $180 + $660 = $9,194 AMD EPYC 9015 59844 ¢ — SabrePC · new · tier 2/sabrepc_ngstate ASRock Rack GENOAD8QM3-2T/BCM 83900 ¢ — Newegg · new · tier 3/newegg_realtime Samsung M321R4GA3BB6-CQK 32GB DDR5-4800 RDIMM 21427 ¢ — SabrePC · new · tier 2/sabrepc_ngstate GIGABYTE AI TOP Radeon AI PRO R9700 32GB (GV-R9700AI TOP-32GD) 171089 ¢ — SabrePC · new · tier 2/sabrepc_ngstate ARCTIC Freezer 4U-SP5 6999 ¢ — ARCTIC · new · tier 1/arctic_jsonld Phanteks Enthoo Pro 2 Server Edition 17999 ¢ — Newegg · new · tier 3/newegg_realtime Seasonic PRIME PX-2200 65999 ¢ — Newegg · new · tier 3/newegg_realtime Listed price only: shipping and sales tax are NOT in this figure. One price per part — the cheapest monitored, non-quarantined observation.
2× Radeon AI PRO R9700 Creator 32GB (R9700 CT) · EPYC 9015 · 8× M321R4GA3BB6-CQK 32GB DDR5-4800 RDIMM 305 GB (derived) DERIVED2× Radeon AI PRO R9700 Creator 32GB (R9700 CT) · EPYC 9015 · 8× M321R4GA3BB6-CQK 32GB DDR5-4800 RDIMM usable memoryArithmetic on observed inputs only. No tunable assumption enters it. available_gb = Σ usable_gb over memory pools GPU VRAM 62.08 + System RAM 243.2 = 305.28 GB GPU VRAM.usable_gb 62.08 GB — 64 GB physical System RAM.usable_gb 243.2 GB — 256 GB physical Usable memory is below physical memory for at least one pool (OS/driver reservation; editable in Settings). usable 640 (observed) OBSERVED2× Radeon AI PRO R9700 Creator 32GB (R9700 CT) · EPYC 9015 · 8× M321R4GA3BB6-CQK 32GB DDR5-4800 RDIMM peak memory bandwidthMeasured or published, passed through unchanged — not computed here. peak memory bandwidth — published at the source, passed through unchanged; nothing is computed here 640 GB/s, as recorded from the GPU VRAM spec of 2× Radeon AI PRO R9700 Creator 32GB (R9700 CT) · EPYC 9015 · 8× M321R4GA3BB6-CQK 32GB DDR5-4800 RDIMM peak_memory_bandwidth 640 GB/s — the GPU VRAM spec of 2× Radeon AI PRO R9700 Creator 32GB (R9700 CT) · EPYC 9015 · 8× M321R4GA3BB6-CQK 32GB DDR5-4800 RDIMM Theoretical peak. Decode reaches a fraction of it — the efficiency factor is in the decode speed's own workings. The fastest pool. A machine with more than one pool places weights fastest-pool-first. Observed, not derived: this app records the figure and its source, and does not re-measure it. GB/s	Usable	9.8 tok/s (estimate) ESTIMATEQwen3-VL 235B-A22B Instruct decode speed on 2× Radeon AI PRO R9700 Creator 32GB (R9700 CT) · EPYC 9015 · 8× M321R4GA3BB6-CQK 32GB DDR5-4800 RDIMMArithmetic that also depends on a tunable assumption, listed below. tok_s = 1 / (Σ_pool bytes_per_token × share_pool / (bandwidth_pool × efficiency_pool) + Σ_pool share_pool × overhead_pool) GPU VRAM: 7.67 GB ÷ (640 GB/s × 0.48) + System RAM: 10.804 GB ÷ (307.2 GB/s × 0.7) + 26.74 ms fixed per-token overhead (share-weighted) → 9.81 tok/s GPU VRAM.bandwidth_gbs 640 GB/s — component spec GPU VRAM.efficiency_moe 0.48 — Settings efficiency.gpu_vram — Discrete GPU: 10 dense + 3 MoE points; fit on published benchmarks, leave-one-out validated — docs/research/12b-estimator-validation.md GPU VRAM.active_share 0.4152 — placement System RAM.bandwidth_gbs 307.2 GB/s — component spec System RAM.efficiency_moe 0.7 — Settings efficiency.server_ddr5 — Server DDR5/MRDIMM single socket: 6 dense + 7 MoE points (small n); fit on published benchmarks, leave-one-out validated — docs/research/12b-estimator-validation.md. The 44.61 ms per-token overhead is a CPU-engine constant, not a property of DDR5: its fit is set by a 128-core Xeon 6980P running llama.cpp (55.3 ms/token on a 3B model) and SGLang (17.0 ms) with the CPU executing every layer (research 12 §6). A GPU spilling into this memory pays it only for the share of each token read from RAM (lib/perf/decode.ts refinement 3). System RAM.active_share 0.5848 — placement per_token_overhead_ms 26.744137424244723 ms — derived: 0.415 × 1.58 ms (GPU VRAM) + 0.585 × 44.61 ms (System RAM) = 26.74 ms bytes_per_token_gb 18.4738 GB MoE: only active parameters are read per token (SPEC §3). Effective bits include embeddings and higher-precision tensors in mixed quants. Assumes active weights are quantized like the file average. KV cache is re-read on every token (DECISIONS D-008). Pools are read sequentially per token, so per-pool times add (harmonic combination). layer split across 2 GPUs: capacity adds, per-token bandwidth does not (layers run sequentially) Per-token overhead on a split placement is DERIVED, not measured: each pool's own single-pool figure is charged in proportion to the share of each token it serves, because the fixed cost is per operation and the split puts that share of the operations on each device. Calibrated with that pool serving the whole model (D-013): GPU VRAM (1.58 ms), System RAM (44.61 ms). Batch size 1 (single user, interactive).	21.1 s (estimate) ESTIMATEQwen3-VL 235B-A22B Instruct time to first token on 2× Radeon AI PRO R9700 Creator 32GB (R9700 CT) · EPYC 9015 · 8× M321R4GA3BB6-CQK 32GB DDR5-4800 RDIMMArithmetic that also depends on a tunable assumption, listed below. ttft_s = prompt_tokens / prefill_tok_s = 32,768 / 1,549.72 = 21.1 s prompt_tokens 32768 tokens prefill_tok_s 1549.72 tok/s No published prefill measurement for this hardware; scaled from peak compute (estimate). Model q_dim unknown — attention FLOPs omitted, so long-prompt TTFT is understated.	$7,462 (derived) DERIVED2× Radeon AI PRO R9700 Creator 32GB (R9700 CT) · EPYC 9015 · 8× M321R4GA3BB6-CQK 32GB DDR5-4800 RDIMM observed priceArithmetic on observed inputs only. No tunable assumption enters it. observed_price = Σ (cheapest observed unit price × quantity) $598 + $839 + $214 × 8 + $1,700 × 2 + $70 + $180 + $660 = $7,462 AMD EPYC 9015 59844 ¢ — SabrePC · new · tier 2/sabrepc_ngstate ASRock Rack GENOAD8QM3-2T/BCM 83900 ¢ — Newegg · new · tier 3/newegg_realtime Samsung M321R4GA3BB6-CQK 32GB DDR5-4800 RDIMM 21427 ¢ — SabrePC · new · tier 2/sabrepc_ngstate ASRock Radeon AI PRO R9700 Creator 32GB (R9700 CT) 169999 ¢ — Newegg · new · tier 2/seed:research ARCTIC Freezer 4U-SP5 6999 ¢ — ARCTIC · new · tier 1/arctic_jsonld Phanteks Enthoo Pro 2 Server Edition 17999 ¢ — Newegg · new · tier 3/newegg_realtime Seasonic PRIME PX-2200 65999 ¢ — Newegg · new · tier 3/newegg_realtime Listed price only: shipping and sales tax are NOT in this figure. One price per part — the cheapest monitored, non-quarantined observation.
2× Radeon AI PRO R9700 32GB (32358-01-20G) · EPYC 9015 · 8× M321R4GA3BB6-CQK 32GB DDR5-4800 RDIMM 305 GB (derived) DERIVED2× Radeon AI PRO R9700 32GB (32358-01-20G) · EPYC 9015 · 8× M321R4GA3BB6-CQK 32GB DDR5-4800 RDIMM usable memoryArithmetic on observed inputs only. No tunable assumption enters it. available_gb = Σ usable_gb over memory pools GPU VRAM 62.08 + System RAM 243.2 = 305.28 GB GPU VRAM.usable_gb 62.08 GB — 64 GB physical System RAM.usable_gb 243.2 GB — 256 GB physical Usable memory is below physical memory for at least one pool (OS/driver reservation; editable in Settings). usable 640 (observed) OBSERVED2× Radeon AI PRO R9700 32GB (32358-01-20G) · EPYC 9015 · 8× M321R4GA3BB6-CQK 32GB DDR5-4800 RDIMM peak memory bandwidthMeasured or published, passed through unchanged — not computed here. peak memory bandwidth — published at the source, passed through unchanged; nothing is computed here 640 GB/s, as recorded from the GPU VRAM spec of 2× Radeon AI PRO R9700 32GB (32358-01-20G) · EPYC 9015 · 8× M321R4GA3BB6-CQK 32GB DDR5-4800 RDIMM peak_memory_bandwidth 640 GB/s — the GPU VRAM spec of 2× Radeon AI PRO R9700 32GB (32358-01-20G) · EPYC 9015 · 8× M321R4GA3BB6-CQK 32GB DDR5-4800 RDIMM Theoretical peak. Decode reaches a fraction of it — the efficiency factor is in the decode speed's own workings. The fastest pool. A machine with more than one pool places weights fastest-pool-first. Observed, not derived: this app records the figure and its source, and does not re-measure it. GB/s	Usable	9.8 tok/s (estimate) ESTIMATEQwen3-VL 235B-A22B Instruct decode speed on 2× Radeon AI PRO R9700 32GB (32358-01-20G) · EPYC 9015 · 8× M321R4GA3BB6-CQK 32GB DDR5-4800 RDIMMArithmetic that also depends on a tunable assumption, listed below. tok_s = 1 / (Σ_pool bytes_per_token × share_pool / (bandwidth_pool × efficiency_pool) + Σ_pool share_pool × overhead_pool) GPU VRAM: 7.67 GB ÷ (640 GB/s × 0.48) + System RAM: 10.804 GB ÷ (307.2 GB/s × 0.7) + 26.74 ms fixed per-token overhead (share-weighted) → 9.81 tok/s GPU VRAM.bandwidth_gbs 640 GB/s — component spec GPU VRAM.efficiency_moe 0.48 — Settings efficiency.gpu_vram — Discrete GPU: 10 dense + 3 MoE points; fit on published benchmarks, leave-one-out validated — docs/research/12b-estimator-validation.md GPU VRAM.active_share 0.4152 — placement System RAM.bandwidth_gbs 307.2 GB/s — component spec System RAM.efficiency_moe 0.7 — Settings efficiency.server_ddr5 — Server DDR5/MRDIMM single socket: 6 dense + 7 MoE points (small n); fit on published benchmarks, leave-one-out validated — docs/research/12b-estimator-validation.md. The 44.61 ms per-token overhead is a CPU-engine constant, not a property of DDR5: its fit is set by a 128-core Xeon 6980P running llama.cpp (55.3 ms/token on a 3B model) and SGLang (17.0 ms) with the CPU executing every layer (research 12 §6). A GPU spilling into this memory pays it only for the share of each token read from RAM (lib/perf/decode.ts refinement 3). System RAM.active_share 0.5848 — placement per_token_overhead_ms 26.744137424244723 ms — derived: 0.415 × 1.58 ms (GPU VRAM) + 0.585 × 44.61 ms (System RAM) = 26.74 ms bytes_per_token_gb 18.4738 GB MoE: only active parameters are read per token (SPEC §3). Effective bits include embeddings and higher-precision tensors in mixed quants. Assumes active weights are quantized like the file average. KV cache is re-read on every token (DECISIONS D-008). Pools are read sequentially per token, so per-pool times add (harmonic combination). layer split across 2 GPUs: capacity adds, per-token bandwidth does not (layers run sequentially) Per-token overhead on a split placement is DERIVED, not measured: each pool's own single-pool figure is charged in proportion to the share of each token it serves, because the fixed cost is per operation and the split puts that share of the operations on each device. Calibrated with that pool serving the whole model (D-013): GPU VRAM (1.58 ms), System RAM (44.61 ms). Batch size 1 (single user, interactive).	21.1 s (estimate) ESTIMATEQwen3-VL 235B-A22B Instruct time to first token on 2× Radeon AI PRO R9700 32GB (32358-01-20G) · EPYC 9015 · 8× M321R4GA3BB6-CQK 32GB DDR5-4800 RDIMMArithmetic that also depends on a tunable assumption, listed below. ttft_s = prompt_tokens / prefill_tok_s = 32,768 / 1,549.72 = 21.1 s prompt_tokens 32768 tokens prefill_tok_s 1549.72 tok/s No published prefill measurement for this hardware; scaled from peak compute (estimate). Model q_dim unknown — attention FLOPs omitted, so long-prompt TTFT is understated.	$7,270 (derived) DERIVED2× Radeon AI PRO R9700 32GB (32358-01-20G) · EPYC 9015 · 8× M321R4GA3BB6-CQK 32GB DDR5-4800 RDIMM observed priceArithmetic on observed inputs only. No tunable assumption enters it. observed_price = Σ (cheapest observed unit price × quantity) $598 + $839 + $214 × 8 + $1,604 × 2 + $70 + $180 + $660 = $7,270 AMD EPYC 9015 59844 ¢ — SabrePC · new · tier 2/sabrepc_ngstate ASRock Rack GENOAD8QM3-2T/BCM 83900 ¢ — Newegg · new · tier 3/newegg_realtime Samsung M321R4GA3BB6-CQK 32GB DDR5-4800 RDIMM 21427 ¢ — SabrePC · new · tier 2/sabrepc_ngstate Sapphire Radeon AI PRO R9700 32GB (32358-01-20G) 160431 ¢ — SabrePC · new · tier 2/sabrepc_ngstate ARCTIC Freezer 4U-SP5 6999 ¢ — ARCTIC · new · tier 1/arctic_jsonld Phanteks Enthoo Pro 2 Server Edition 17999 ¢ — Newegg · new · tier 3/newegg_realtime Seasonic PRIME PX-2200 65999 ¢ — Newegg · new · tier 3/newegg_realtime Listed price only: shipping and sales tax are NOT in this figure. One price per part — the cheapest monitored, non-quarantined observation.
2× AI TOP Radeon AI PRO R9700 32GB (GV-R9700AI TOP-32GD) · EPYC 9015 · 8× M321R4GA3BB6-CQK 32GB DDR5-4800 RDIMM 305 GB (derived) DERIVED2× AI TOP Radeon AI PRO R9700 32GB (GV-R9700AI TOP-32GD) · EPYC 9015 · 8× M321R4GA3BB6-CQK 32GB DDR5-4800 RDIMM usable memoryArithmetic on observed inputs only. No tunable assumption enters it. available_gb = Σ usable_gb over memory pools GPU VRAM 62.08 + System RAM 243.2 = 305.28 GB GPU VRAM.usable_gb 62.08 GB — 64 GB physical System RAM.usable_gb 243.2 GB — 256 GB physical Usable memory is below physical memory for at least one pool (OS/driver reservation; editable in Settings). usable 640 (observed) OBSERVED2× AI TOP Radeon AI PRO R9700 32GB (GV-R9700AI TOP-32GD) · EPYC 9015 · 8× M321R4GA3BB6-CQK 32GB DDR5-4800 RDIMM peak memory bandwidthMeasured or published, passed through unchanged — not computed here. peak memory bandwidth — published at the source, passed through unchanged; nothing is computed here 640 GB/s, as recorded from the GPU VRAM spec of 2× AI TOP Radeon AI PRO R9700 32GB (GV-R9700AI TOP-32GD) · EPYC 9015 · 8× M321R4GA3BB6-CQK 32GB DDR5-4800 RDIMM peak_memory_bandwidth 640 GB/s — the GPU VRAM spec of 2× AI TOP Radeon AI PRO R9700 32GB (GV-R9700AI TOP-32GD) · EPYC 9015 · 8× M321R4GA3BB6-CQK 32GB DDR5-4800 RDIMM Theoretical peak. Decode reaches a fraction of it — the efficiency factor is in the decode speed's own workings. The fastest pool. A machine with more than one pool places weights fastest-pool-first. Observed, not derived: this app records the figure and its source, and does not re-measure it. GB/s	Usable	9.8 tok/s (estimate) ESTIMATEQwen3-VL 235B-A22B Instruct decode speed on 2× AI TOP Radeon AI PRO R9700 32GB (GV-R9700AI TOP-32GD) · EPYC 9015 · 8× M321R4GA3BB6-CQK 32GB DDR5-4800 RDIMMArithmetic that also depends on a tunable assumption, listed below. tok_s = 1 / (Σ_pool bytes_per_token × share_pool / (bandwidth_pool × efficiency_pool) + Σ_pool share_pool × overhead_pool) GPU VRAM: 7.67 GB ÷ (640 GB/s × 0.48) + System RAM: 10.804 GB ÷ (307.2 GB/s × 0.7) + 26.74 ms fixed per-token overhead (share-weighted) → 9.81 tok/s GPU VRAM.bandwidth_gbs 640 GB/s — component spec GPU VRAM.efficiency_moe 0.48 — Settings efficiency.gpu_vram — Discrete GPU: 10 dense + 3 MoE points; fit on published benchmarks, leave-one-out validated — docs/research/12b-estimator-validation.md GPU VRAM.active_share 0.4152 — placement System RAM.bandwidth_gbs 307.2 GB/s — component spec System RAM.efficiency_moe 0.7 — Settings efficiency.server_ddr5 — Server DDR5/MRDIMM single socket: 6 dense + 7 MoE points (small n); fit on published benchmarks, leave-one-out validated — docs/research/12b-estimator-validation.md. The 44.61 ms per-token overhead is a CPU-engine constant, not a property of DDR5: its fit is set by a 128-core Xeon 6980P running llama.cpp (55.3 ms/token on a 3B model) and SGLang (17.0 ms) with the CPU executing every layer (research 12 §6). A GPU spilling into this memory pays it only for the share of each token read from RAM (lib/perf/decode.ts refinement 3). System RAM.active_share 0.5848 — placement per_token_overhead_ms 26.744137424244723 ms — derived: 0.415 × 1.58 ms (GPU VRAM) + 0.585 × 44.61 ms (System RAM) = 26.74 ms bytes_per_token_gb 18.4738 GB MoE: only active parameters are read per token (SPEC §3). Effective bits include embeddings and higher-precision tensors in mixed quants. Assumes active weights are quantized like the file average. KV cache is re-read on every token (DECISIONS D-008). Pools are read sequentially per token, so per-pool times add (harmonic combination). layer split across 2 GPUs: capacity adds, per-token bandwidth does not (layers run sequentially) Per-token overhead on a split placement is DERIVED, not measured: each pool's own single-pool figure is charged in proportion to the share of each token it serves, because the fixed cost is per operation and the split puts that share of the operations on each device. Calibrated with that pool serving the whole model (D-013): GPU VRAM (1.58 ms), System RAM (44.61 ms). Batch size 1 (single user, interactive).	21.1 s (estimate) ESTIMATEQwen3-VL 235B-A22B Instruct time to first token on 2× AI TOP Radeon AI PRO R9700 32GB (GV-R9700AI TOP-32GD) · EPYC 9015 · 8× M321R4GA3BB6-CQK 32GB DDR5-4800 RDIMMArithmetic that also depends on a tunable assumption, listed below. ttft_s = prompt_tokens / prefill_tok_s = 32,768 / 1,549.72 = 21.1 s prompt_tokens 32768 tokens prefill_tok_s 1549.72 tok/s No published prefill measurement for this hardware; scaled from peak compute (estimate). Model q_dim unknown — attention FLOPs omitted, so long-prompt TTFT is understated.	$7,483 (derived) DERIVED2× AI TOP Radeon AI PRO R9700 32GB (GV-R9700AI TOP-32GD) · EPYC 9015 · 8× M321R4GA3BB6-CQK 32GB DDR5-4800 RDIMM observed priceArithmetic on observed inputs only. No tunable assumption enters it. observed_price = Σ (cheapest observed unit price × quantity) $598 + $839 + $214 × 8 + $1,711 × 2 + $70 + $180 + $660 = $7,483 AMD EPYC 9015 59844 ¢ — SabrePC · new · tier 2/sabrepc_ngstate ASRock Rack GENOAD8QM3-2T/BCM 83900 ¢ — Newegg · new · tier 3/newegg_realtime Samsung M321R4GA3BB6-CQK 32GB DDR5-4800 RDIMM 21427 ¢ — SabrePC · new · tier 2/sabrepc_ngstate GIGABYTE AI TOP Radeon AI PRO R9700 32GB (GV-R9700AI TOP-32GD) 171089 ¢ — SabrePC · new · tier 2/sabrepc_ngstate ARCTIC Freezer 4U-SP5 6999 ¢ — ARCTIC · new · tier 1/arctic_jsonld Phanteks Enthoo Pro 2 Server Edition 17999 ¢ — Newegg · new · tier 3/newegg_realtime Seasonic PRIME PX-2200 65999 ¢ — Newegg · new · tier 3/newegg_realtime Listed price only: shipping and sales tax are NOT in this figure. One price per part — the cheapest monitored, non-quarantined observation.
3× GeForce RTX 5080 WINDFORCE OC 16G (GV-N5080WF3OC-16GD) · EPYC 9015 · 8× M321R4GA3BB6-CQK 32GB DDR5-4800 RDIMM 290 GB (derived) DERIVED3× GeForce RTX 5080 WINDFORCE OC 16G (GV-N5080WF3OC-16GD) · EPYC 9015 · 8× M321R4GA3BB6-CQK 32GB DDR5-4800 RDIMM usable memoryArithmetic on observed inputs only. No tunable assumption enters it. available_gb = Σ usable_gb over memory pools GPU VRAM 46.56 + System RAM 243.2 = 289.76 GB GPU VRAM.usable_gb 46.56 GB — 48 GB physical System RAM.usable_gb 243.2 GB — 256 GB physical Usable memory is below physical memory for at least one pool (OS/driver reservation; editable in Settings). usable 960 (observed) OBSERVED3× GeForce RTX 5080 WINDFORCE OC 16G (GV-N5080WF3OC-16GD) · EPYC 9015 · 8× M321R4GA3BB6-CQK 32GB DDR5-4800 RDIMM peak memory bandwidthMeasured or published, passed through unchanged — not computed here. peak memory bandwidth — published at the source, passed through unchanged; nothing is computed here 960 GB/s, as recorded from the GPU VRAM spec of 3× GeForce RTX 5080 WINDFORCE OC 16G (GV-N5080WF3OC-16GD) · EPYC 9015 · 8× M321R4GA3BB6-CQK 32GB DDR5-4800 RDIMM peak_memory_bandwidth 960 GB/s — the GPU VRAM spec of 3× GeForce RTX 5080 WINDFORCE OC 16G (GV-N5080WF3OC-16GD) · EPYC 9015 · 8× M321R4GA3BB6-CQK 32GB DDR5-4800 RDIMM Theoretical peak. Decode reaches a fraction of it — the efficiency factor is in the decode speed's own workings. The fastest pool. A machine with more than one pool places weights fastest-pool-first. Observed, not derived: this app records the figure and its source, and does not re-measure it. GB/s	Usable	9.7 tok/s (estimate) ESTIMATEQwen3-VL 235B-A22B Instruct decode speed on 3× GeForce RTX 5080 WINDFORCE OC 16G (GV-N5080WF3OC-16GD) · EPYC 9015 · 8× M321R4GA3BB6-CQK 32GB DDR5-4800 RDIMMArithmetic that also depends on a tunable assumption, listed below. tok_s = 1 / (Σ_pool bytes_per_token × share_pool / (bandwidth_pool × efficiency_pool) + Σ_pool share_pool × overhead_pool) GPU VRAM: 5.753 GB ÷ (960 GB/s × 0.48) + System RAM: 12.721 GB ÷ (307.2 GB/s × 0.7) + 31.21 ms fixed per-token overhead (share-weighted) → 9.72 tok/s GPU VRAM.bandwidth_gbs 960 GB/s — component spec GPU VRAM.efficiency_moe 0.48 — Settings efficiency.gpu_vram — Discrete GPU: 10 dense + 3 MoE points; fit on published benchmarks, leave-one-out validated — docs/research/12b-estimator-validation.md GPU VRAM.active_share 0.3114 — placement System RAM.bandwidth_gbs 307.2 GB/s — component spec System RAM.efficiency_moe 0.7 — Settings efficiency.server_ddr5 — Server DDR5/MRDIMM single socket: 6 dense + 7 MoE points (small n); fit on published benchmarks, leave-one-out validated — docs/research/12b-estimator-validation.md. The 44.61 ms per-token overhead is a CPU-engine constant, not a property of DDR5: its fit is set by a 128-core Xeon 6980P running llama.cpp (55.3 ms/token on a 3B model) and SGLang (17.0 ms) with the CPU executing every layer (research 12 §6). A GPU spilling into this memory pays it only for the share of each token read from RAM (lib/perf/decode.ts refinement 3). System RAM.active_share 0.6886 — placement per_token_overhead_ms 31.210603068183538 ms — derived: 0.311 × 1.58 ms (GPU VRAM) + 0.689 × 44.61 ms (System RAM) = 31.21 ms bytes_per_token_gb 18.4738 GB MoE: only active parameters are read per token (SPEC §3). Effective bits include embeddings and higher-precision tensors in mixed quants. Assumes active weights are quantized like the file average. KV cache is re-read on every token (DECISIONS D-008). Pools are read sequentially per token, so per-pool times add (harmonic combination). layer split across 3 GPUs: capacity adds, per-token bandwidth does not (layers run sequentially) Per-token overhead on a split placement is DERIVED, not measured: each pool's own single-pool figure is charged in proportion to the share of each token it serves, because the fixed cost is per operation and the split puts that share of the operations on each device. Calibrated with that pool serving the whole model (D-013): GPU VRAM (1.58 ms), System RAM (44.61 ms). Batch size 1 (single user, interactive).	35.9 s (estimate) ESTIMATEQwen3-VL 235B-A22B Instruct time to first token on 3× GeForce RTX 5080 WINDFORCE OC 16G (GV-N5080WF3OC-16GD) · EPYC 9015 · 8× M321R4GA3BB6-CQK 32GB DDR5-4800 RDIMMArithmetic that also depends on a tunable assumption, listed below. ttft_s = prompt_tokens / prefill_tok_s = 32,768 / 913.6 = 35.9 s prompt_tokens 32768 tokens prefill_tok_s 913.6 tok/s No published prefill measurement for this hardware; scaled from peak compute (estimate). Model q_dim unknown — attention FLOPs omitted, so long-prompt TTFT is understated.	$9,162 (derived) DERIVED3× GeForce RTX 5080 WINDFORCE OC 16G (GV-N5080WF3OC-16GD) · EPYC 9015 · 8× M321R4GA3BB6-CQK 32GB DDR5-4800 RDIMM observed priceArithmetic on observed inputs only. No tunable assumption enters it. observed_price = Σ (cheapest observed unit price × quantity) $598 + $839 + $214 × 8 + $1,700 × 3 + $70 + $180 + $660 = $9,162 AMD EPYC 9015 59844 ¢ — SabrePC · new · tier 2/sabrepc_ngstate ASRock Rack GENOAD8QM3-2T/BCM 83900 ¢ — Newegg · new · tier 3/newegg_realtime Samsung M321R4GA3BB6-CQK 32GB DDR5-4800 RDIMM 21427 ¢ — SabrePC · new · tier 2/sabrepc_ngstate GIGABYTE GeForce RTX 5080 WINDFORCE OC 16G (GV-N5080WF3OC-16GD) 169999 ¢ — Newegg · new · tier 3/newegg_realtime ARCTIC Freezer 4U-SP5 6999 ¢ — ARCTIC · new · tier 1/arctic_jsonld Phanteks Enthoo Pro 2 Server Edition 17999 ¢ — Newegg · new · tier 3/newegg_realtime Seasonic PRIME PX-2200 65999 ¢ — Newegg · new · tier 3/newegg_realtime Listed price only: shipping and sales tax are NOT in this figure. One price per part — the cheapest monitored, non-quarantined observation.
3× GeForce RTX 5080 16G SHADOW 3X OC · EPYC 9015 · 8× M321R4GA3BB6-CQK 32GB DDR5-4800 RDIMM 290 GB (derived) DERIVED3× GeForce RTX 5080 16G SHADOW 3X OC · EPYC 9015 · 8× M321R4GA3BB6-CQK 32GB DDR5-4800 RDIMM usable memoryArithmetic on observed inputs only. No tunable assumption enters it. available_gb = Σ usable_gb over memory pools GPU VRAM 46.56 + System RAM 243.2 = 289.76 GB GPU VRAM.usable_gb 46.56 GB — 48 GB physical System RAM.usable_gb 243.2 GB — 256 GB physical Usable memory is below physical memory for at least one pool (OS/driver reservation; editable in Settings). usable 960 (observed) OBSERVED3× GeForce RTX 5080 16G SHADOW 3X OC · EPYC 9015 · 8× M321R4GA3BB6-CQK 32GB DDR5-4800 RDIMM peak memory bandwidthMeasured or published, passed through unchanged — not computed here. peak memory bandwidth — published at the source, passed through unchanged; nothing is computed here 960 GB/s, as recorded from the GPU VRAM spec of 3× GeForce RTX 5080 16G SHADOW 3X OC · EPYC 9015 · 8× M321R4GA3BB6-CQK 32GB DDR5-4800 RDIMM peak_memory_bandwidth 960 GB/s — the GPU VRAM spec of 3× GeForce RTX 5080 16G SHADOW 3X OC · EPYC 9015 · 8× M321R4GA3BB6-CQK 32GB DDR5-4800 RDIMM Theoretical peak. Decode reaches a fraction of it — the efficiency factor is in the decode speed's own workings. The fastest pool. A machine with more than one pool places weights fastest-pool-first. Observed, not derived: this app records the figure and its source, and does not re-measure it. GB/s	Usable	9.7 tok/s (estimate) ESTIMATEQwen3-VL 235B-A22B Instruct decode speed on 3× GeForce RTX 5080 16G SHADOW 3X OC · EPYC 9015 · 8× M321R4GA3BB6-CQK 32GB DDR5-4800 RDIMMArithmetic that also depends on a tunable assumption, listed below. tok_s = 1 / (Σ_pool bytes_per_token × share_pool / (bandwidth_pool × efficiency_pool) + Σ_pool share_pool × overhead_pool) GPU VRAM: 5.753 GB ÷ (960 GB/s × 0.48) + System RAM: 12.721 GB ÷ (307.2 GB/s × 0.7) + 31.21 ms fixed per-token overhead (share-weighted) → 9.72 tok/s GPU VRAM.bandwidth_gbs 960 GB/s — component spec GPU VRAM.efficiency_moe 0.48 — Settings efficiency.gpu_vram — Discrete GPU: 10 dense + 3 MoE points; fit on published benchmarks, leave-one-out validated — docs/research/12b-estimator-validation.md GPU VRAM.active_share 0.3114 — placement System RAM.bandwidth_gbs 307.2 GB/s — component spec System RAM.efficiency_moe 0.7 — Settings efficiency.server_ddr5 — Server DDR5/MRDIMM single socket: 6 dense + 7 MoE points (small n); fit on published benchmarks, leave-one-out validated — docs/research/12b-estimator-validation.md. The 44.61 ms per-token overhead is a CPU-engine constant, not a property of DDR5: its fit is set by a 128-core Xeon 6980P running llama.cpp (55.3 ms/token on a 3B model) and SGLang (17.0 ms) with the CPU executing every layer (research 12 §6). A GPU spilling into this memory pays it only for the share of each token read from RAM (lib/perf/decode.ts refinement 3). System RAM.active_share 0.6886 — placement per_token_overhead_ms 31.210603068183538 ms — derived: 0.311 × 1.58 ms (GPU VRAM) + 0.689 × 44.61 ms (System RAM) = 31.21 ms bytes_per_token_gb 18.4738 GB MoE: only active parameters are read per token (SPEC §3). Effective bits include embeddings and higher-precision tensors in mixed quants. Assumes active weights are quantized like the file average. KV cache is re-read on every token (DECISIONS D-008). Pools are read sequentially per token, so per-pool times add (harmonic combination). layer split across 3 GPUs: capacity adds, per-token bandwidth does not (layers run sequentially) Per-token overhead on a split placement is DERIVED, not measured: each pool's own single-pool figure is charged in proportion to the share of each token it serves, because the fixed cost is per operation and the split puts that share of the operations on each device. Calibrated with that pool serving the whole model (D-013): GPU VRAM (1.58 ms), System RAM (44.61 ms). Batch size 1 (single user, interactive).	35.9 s (estimate) ESTIMATEQwen3-VL 235B-A22B Instruct time to first token on 3× GeForce RTX 5080 16G SHADOW 3X OC · EPYC 9015 · 8× M321R4GA3BB6-CQK 32GB DDR5-4800 RDIMMArithmetic that also depends on a tunable assumption, listed below. ttft_s = prompt_tokens / prefill_tok_s = 32,768 / 913.6 = 35.9 s prompt_tokens 32768 tokens prefill_tok_s 913.6 tok/s No published prefill measurement for this hardware; scaled from peak compute (estimate). Model q_dim unknown — attention FLOPs omitted, so long-prompt TTFT is understated.	$9,132 (derived) DERIVED3× GeForce RTX 5080 16G SHADOW 3X OC · EPYC 9015 · 8× M321R4GA3BB6-CQK 32GB DDR5-4800 RDIMM observed priceArithmetic on observed inputs only. No tunable assumption enters it. observed_price = Σ (cheapest observed unit price × quantity) $598 + $839 + $214 × 8 + $1,690 × 3 + $70 + $180 + $660 = $9,132 AMD EPYC 9015 59844 ¢ — SabrePC · new · tier 2/sabrepc_ngstate ASRock Rack GENOAD8QM3-2T/BCM 83900 ¢ — Newegg · new · tier 3/newegg_realtime Samsung M321R4GA3BB6-CQK 32GB DDR5-4800 RDIMM 21427 ¢ — SabrePC · new · tier 2/sabrepc_ngstate MSI GeForce RTX 5080 16G SHADOW 3X OC 168999 ¢ — Newegg · new · tier 3/newegg_realtime ARCTIC Freezer 4U-SP5 6999 ¢ — ARCTIC · new · tier 1/arctic_jsonld Phanteks Enthoo Pro 2 Server Edition 17999 ¢ — Newegg · new · tier 3/newegg_realtime Seasonic PRIME PX-2200 65999 ¢ — Newegg · new · tier 3/newegg_realtime Listed price only: shipping and sales tax are NOT in this figure. One price per part — the cheapest monitored, non-quarantined observation.
1× GeForce RTX 5090 32G GAMING TRIO OC · EPYC 9015 · 8× M321R4GA3BB6-CQK 32GB DDR5-4800 RDIMM 274 GB (derived) DERIVED1× GeForce RTX 5090 32G GAMING TRIO OC · EPYC 9015 · 8× M321R4GA3BB6-CQK 32GB DDR5-4800 RDIMM usable memoryArithmetic on observed inputs only. No tunable assumption enters it. available_gb = Σ usable_gb over memory pools GPU VRAM 31.04 + System RAM 243.2 = 274.24 GB GPU VRAM.usable_gb 31.04 GB — 32 GB physical System RAM.usable_gb 243.2 GB — 256 GB physical Usable memory is below physical memory for at least one pool (OS/driver reservation; editable in Settings). usable 1,792 (observed) OBSERVED1× GeForce RTX 5090 32G GAMING TRIO OC · EPYC 9015 · 8× M321R4GA3BB6-CQK 32GB DDR5-4800 RDIMM peak memory bandwidthMeasured or published, passed through unchanged — not computed here. peak memory bandwidth — published at the source, passed through unchanged; nothing is computed here 1,792 GB/s, as recorded from the GPU VRAM spec of 1× GeForce RTX 5090 32G GAMING TRIO OC · EPYC 9015 · 8× M321R4GA3BB6-CQK 32GB DDR5-4800 RDIMM peak_memory_bandwidth 1,792 GB/s — the GPU VRAM spec of 1× GeForce RTX 5090 32G GAMING TRIO OC · EPYC 9015 · 8× M321R4GA3BB6-CQK 32GB DDR5-4800 RDIMM Theoretical peak. Decode reaches a fraction of it — the efficiency factor is in the decode speed's own workings. The fastest pool. A machine with more than one pool places weights fastest-pool-first. Observed, not derived: this app records the figure and its source, and does not re-measure it. GB/s	Usable	9.2 tok/s (estimate) ESTIMATEQwen3-VL 235B-A22B Instruct decode speed on 1× GeForce RTX 5090 32G GAMING TRIO OC · EPYC 9015 · 8× M321R4GA3BB6-CQK 32GB DDR5-4800 RDIMMArithmetic that also depends on a tunable assumption, listed below. tok_s = 1 / (Σ_pool bytes_per_token × share_pool / (bandwidth_pool × efficiency_pool) + Σ_pool share_pool × overhead_pool) GPU VRAM: 3.835 GB ÷ (1,792 GB/s × 0.48) + System RAM: 14.639 GB ÷ (307.2 GB/s × 0.7) + 35.68 ms fixed per-token overhead (share-weighted) → 9.24 tok/s GPU VRAM.bandwidth_gbs 1792 GB/s — component spec GPU VRAM.efficiency_moe 0.48 — Settings efficiency.gpu_vram — Discrete GPU: 10 dense + 3 MoE points; fit on published benchmarks, leave-one-out validated — docs/research/12b-estimator-validation.md GPU VRAM.active_share 0.2076 — placement System RAM.bandwidth_gbs 307.2 GB/s — component spec System RAM.efficiency_moe 0.7 — Settings efficiency.server_ddr5 — Server DDR5/MRDIMM single socket: 6 dense + 7 MoE points (small n); fit on published benchmarks, leave-one-out validated — docs/research/12b-estimator-validation.md. The 44.61 ms per-token overhead is a CPU-engine constant, not a property of DDR5: its fit is set by a 128-core Xeon 6980P running llama.cpp (55.3 ms/token on a 3B model) and SGLang (17.0 ms) with the CPU executing every layer (research 12 §6). A GPU spilling into this memory pays it only for the share of each token read from RAM (lib/perf/decode.ts refinement 3). System RAM.active_share 0.7924 — placement per_token_overhead_ms 35.67706871212236 ms — derived: 0.208 × 1.58 ms (GPU VRAM) + 0.792 × 44.61 ms (System RAM) = 35.68 ms bytes_per_token_gb 18.4738 GB MoE: only active parameters are read per token (SPEC §3). Effective bits include embeddings and higher-precision tensors in mixed quants. Assumes active weights are quantized like the file average. KV cache is re-read on every token (DECISIONS D-008). Pools are read sequentially per token, so per-pool times add (harmonic combination). Per-token overhead on a split placement is DERIVED, not measured: each pool's own single-pool figure is charged in proportion to the share of each token it serves, because the fixed cost is per operation and the split puts that share of the operations on each device. Calibrated with that pool serving the whole model (D-013): GPU VRAM (1.58 ms), System RAM (44.61 ms). Batch size 1 (single user, interactive).	19.3 s (estimate) ESTIMATEQwen3-VL 235B-A22B Instruct time to first token on 1× GeForce RTX 5090 32G GAMING TRIO OC · EPYC 9015 · 8× M321R4GA3BB6-CQK 32GB DDR5-4800 RDIMMArithmetic that also depends on a tunable assumption, listed below. ttft_s = prompt_tokens / prefill_tok_s = 32,768 / 1,699.82 = 19.3 s prompt_tokens 32768 tokens prefill_tok_s 1699.82 tok/s No published prefill measurement for this hardware; scaled from peak compute (estimate). Model q_dim unknown — attention FLOPs omitted, so long-prompt TTFT is understated.	$9,162 (derived) DERIVED1× GeForce RTX 5090 32G GAMING TRIO OC · EPYC 9015 · 8× M321R4GA3BB6-CQK 32GB DDR5-4800 RDIMM observed priceArithmetic on observed inputs only. No tunable assumption enters it. observed_price = Σ (cheapest observed unit price × quantity) $598 + $839 + $214 × 8 + $5,100 + $70 + $180 + $660 = $9,162 AMD EPYC 9015 59844 ¢ — SabrePC · new · tier 2/sabrepc_ngstate ASRock Rack GENOAD8QM3-2T/BCM 83900 ¢ — Newegg · new · tier 3/newegg_realtime Samsung M321R4GA3BB6-CQK 32GB DDR5-4800 RDIMM 21427 ¢ — SabrePC · new · tier 2/sabrepc_ngstate MSI GeForce RTX 5090 32G GAMING TRIO OC 509999 ¢ — Newegg · new · tier 3/newegg_realtime ARCTIC Freezer 4U-SP5 6999 ¢ — ARCTIC · new · tier 1/arctic_jsonld Phanteks Enthoo Pro 2 Server Edition 17999 ¢ — Newegg · new · tier 3/newegg_realtime Seasonic PRIME PX-2200 65999 ¢ — Newegg · new · tier 3/newegg_realtime Listed price only: shipping and sales tax are NOT in this figure. One price per part — the cheapest monitored, non-quarantined observation.
Hybrid · GeForce RTX 5090 32G GAMING TRIO OC + EPYC 9015 · 8× M321R8GA0BB0-CQK 64GB DDR5-4800 RDIMM 517 GB (derived) DERIVEDHybrid · GeForce RTX 5090 32G GAMING TRIO OC + EPYC 9015 · 8× M321R8GA0BB0-CQK 64GB DDR5-4800 RDIMM usable memoryArithmetic on observed inputs only. No tunable assumption enters it. available_gb = Σ usable_gb over memory pools GPU VRAM 31.04 + System RAM 486.4 = 517.44 GB GPU VRAM.usable_gb 31.04 GB — 32 GB physical System RAM.usable_gb 486.4 GB — 512 GB physical Usable memory is below physical memory for at least one pool (OS/driver reservation; editable in Settings). usable 1,792 (observed) OBSERVEDHybrid · GeForce RTX 5090 32G GAMING TRIO OC + EPYC 9015 · 8× M321R8GA0BB0-CQK 64GB DDR5-4800 RDIMM peak memory bandwidthMeasured or published, passed through unchanged — not computed here. peak memory bandwidth — published at the source, passed through unchanged; nothing is computed here 1,792 GB/s, as recorded from the GPU VRAM spec of Hybrid · GeForce RTX 5090 32G GAMING TRIO OC + EPYC 9015 · 8× M321R8GA0BB0-CQK 64GB DDR5-4800 RDIMM peak_memory_bandwidth 1,792 GB/s — the GPU VRAM spec of Hybrid · GeForce RTX 5090 32G GAMING TRIO OC + EPYC 9015 · 8× M321R8GA0BB0-CQK 64GB DDR5-4800 RDIMM Theoretical peak. Decode reaches a fraction of it — the efficiency factor is in the decode speed's own workings. The fastest pool. A machine with more than one pool places weights fastest-pool-first. Observed, not derived: this app records the figure and its source, and does not re-measure it. GB/s	Usable	9.2 tok/s (estimate) ESTIMATEQwen3-VL 235B-A22B Instruct decode speed on Hybrid · GeForce RTX 5090 32G GAMING TRIO OC + EPYC 9015 · 8× M321R8GA0BB0-CQK 64GB DDR5-4800 RDIMMArithmetic that also depends on a tunable assumption, listed below. tok_s = 1 / (Σ_pool bytes_per_token × share_pool / (bandwidth_pool × efficiency_pool) + Σ_pool share_pool × overhead_pool) GPU VRAM: 3.835 GB ÷ (1,792 GB/s × 0.48) + System RAM: 14.639 GB ÷ (307.2 GB/s × 0.7) + 35.68 ms fixed per-token overhead (share-weighted) → 9.24 tok/s GPU VRAM.bandwidth_gbs 1792 GB/s — component spec GPU VRAM.efficiency_moe 0.48 — Settings efficiency.gpu_vram — Discrete GPU: 10 dense + 3 MoE points; fit on published benchmarks, leave-one-out validated — docs/research/12b-estimator-validation.md GPU VRAM.active_share 0.2076 — placement System RAM.bandwidth_gbs 307.2 GB/s — component spec System RAM.efficiency_moe 0.7 — Settings efficiency.server_ddr5 — Server DDR5/MRDIMM single socket: 6 dense + 7 MoE points (small n); fit on published benchmarks, leave-one-out validated — docs/research/12b-estimator-validation.md. The 44.61 ms per-token overhead is a CPU-engine constant, not a property of DDR5: its fit is set by a 128-core Xeon 6980P running llama.cpp (55.3 ms/token on a 3B model) and SGLang (17.0 ms) with the CPU executing every layer (research 12 §6). A GPU spilling into this memory pays it only for the share of each token read from RAM (lib/perf/decode.ts refinement 3). System RAM.active_share 0.7924 — placement per_token_overhead_ms 35.67706871212236 ms — derived: 0.208 × 1.58 ms (GPU VRAM) + 0.792 × 44.61 ms (System RAM) = 35.68 ms bytes_per_token_gb 18.4738 GB MoE: only active parameters are read per token (SPEC §3). Effective bits include embeddings and higher-precision tensors in mixed quants. Assumes active weights are quantized like the file average. KV cache is re-read on every token (DECISIONS D-008). Pools are read sequentially per token, so per-pool times add (harmonic combination). Per-token overhead on a split placement is DERIVED, not measured: each pool's own single-pool figure is charged in proportion to the share of each token it serves, because the fixed cost is per operation and the split puts that share of the operations on each device. Calibrated with that pool serving the whole model (D-013): GPU VRAM (1.58 ms), System RAM (44.61 ms). Batch size 1 (single user, interactive).	19.3 s (estimate) ESTIMATEQwen3-VL 235B-A22B Instruct time to first token on Hybrid · GeForce RTX 5090 32G GAMING TRIO OC + EPYC 9015 · 8× M321R8GA0BB0-CQK 64GB DDR5-4800 RDIMMArithmetic that also depends on a tunable assumption, listed below. ttft_s = prompt_tokens / prefill_tok_s = 32,768 / 1,699.82 = 19.3 s prompt_tokens 32768 tokens prefill_tok_s 1699.82 tok/s No published prefill measurement for this hardware; scaled from peak compute (estimate). Model q_dim unknown — attention FLOPs omitted, so long-prompt TTFT is understated.	$9,824 (derived) DERIVEDHybrid · GeForce RTX 5090 32G GAMING TRIO OC + EPYC 9015 · 8× M321R8GA0BB0-CQK 64GB DDR5-4800 RDIMM observed priceArithmetic on observed inputs only. No tunable assumption enters it. observed_price = Σ (cheapest observed unit price × quantity) $598 + $839 + $297 × 8 + $5,100 + $70 + $180 + $660 = $9,824 AMD EPYC 9015 59844 ¢ — SabrePC · new · tier 2/sabrepc_ngstate ASRock Rack GENOAD8QM3-2T/BCM 83900 ¢ — Newegg · new · tier 3/newegg_realtime Samsung M321R8GA0BB0-CQK 64GB DDR5-4800 RDIMM 29707 ¢ — SabrePC · new · tier 2/sabrepc_ngstate MSI GeForce RTX 5090 32G GAMING TRIO OC 509999 ¢ — Newegg · new · tier 3/newegg_realtime ARCTIC Freezer 4U-SP5 6999 ¢ — ARCTIC · new · tier 1/arctic_jsonld Phanteks Enthoo Pro 2 Server Edition 17999 ¢ — Newegg · new · tier 3/newegg_realtime Seasonic PRIME PX-2200 65999 ¢ — Newegg · new · tier 3/newegg_realtime Listed price only: shipping and sales tax are NOT in this figure. One price per part — the cheapest monitored, non-quarantined observation.
Hybrid · GeForce RTX 5090 32G GAMING TRIO OC + EPYC 9015 · 8× Server Premier KSM64R52BD4-96MC 96GB DDR5-6400 RDIMM 761 GB (derived) DERIVEDHybrid · GeForce RTX 5090 32G GAMING TRIO OC + EPYC 9015 · 8× Server Premier KSM64R52BD4-96MC 96GB DDR5-6400 RDIMM usable memoryArithmetic on observed inputs only. No tunable assumption enters it. available_gb = Σ usable_gb over memory pools GPU VRAM 31.04 + System RAM 729.6 = 760.64 GB GPU VRAM.usable_gb 31.04 GB — 32 GB physical System RAM.usable_gb 729.5999999999999 GB — 768 GB physical Usable memory is below physical memory for at least one pool (OS/driver reservation; editable in Settings). usable 1,792 (observed) OBSERVEDHybrid · GeForce RTX 5090 32G GAMING TRIO OC + EPYC 9015 · 8× Server Premier KSM64R52BD4-96MC 96GB DDR5-6400 RDIMM peak memory bandwidthMeasured or published, passed through unchanged — not computed here. peak memory bandwidth — published at the source, passed through unchanged; nothing is computed here 1,792 GB/s, as recorded from the GPU VRAM spec of Hybrid · GeForce RTX 5090 32G GAMING TRIO OC + EPYC 9015 · 8× Server Premier KSM64R52BD4-96MC 96GB DDR5-6400 RDIMM peak_memory_bandwidth 1,792 GB/s — the GPU VRAM spec of Hybrid · GeForce RTX 5090 32G GAMING TRIO OC + EPYC 9015 · 8× Server Premier KSM64R52BD4-96MC 96GB DDR5-6400 RDIMM Theoretical peak. Decode reaches a fraction of it — the efficiency factor is in the decode speed's own workings. The fastest pool. A machine with more than one pool places weights fastest-pool-first. Observed, not derived: this app records the figure and its source, and does not re-measure it. GB/s	Usable	9.2 tok/s (estimate) ESTIMATEQwen3-VL 235B-A22B Instruct decode speed on Hybrid · GeForce RTX 5090 32G GAMING TRIO OC + EPYC 9015 · 8× Server Premier KSM64R52BD4-96MC 96GB DDR5-6400 RDIMMArithmetic that also depends on a tunable assumption, listed below. tok_s = 1 / (Σ_pool bytes_per_token × share_pool / (bandwidth_pool × efficiency_pool) + Σ_pool share_pool × overhead_pool) GPU VRAM: 3.835 GB ÷ (1,792 GB/s × 0.48) + System RAM: 14.639 GB ÷ (307.2 GB/s × 0.7) + 35.68 ms fixed per-token overhead (share-weighted) → 9.24 tok/s GPU VRAM.bandwidth_gbs 1792 GB/s — component spec GPU VRAM.efficiency_moe 0.48 — Settings efficiency.gpu_vram — Discrete GPU: 10 dense + 3 MoE points; fit on published benchmarks, leave-one-out validated — docs/research/12b-estimator-validation.md GPU VRAM.active_share 0.2076 — placement System RAM.bandwidth_gbs 307.2 GB/s — component spec System RAM.efficiency_moe 0.7 — Settings efficiency.server_ddr5 — Server DDR5/MRDIMM single socket: 6 dense + 7 MoE points (small n); fit on published benchmarks, leave-one-out validated — docs/research/12b-estimator-validation.md. The 44.61 ms per-token overhead is a CPU-engine constant, not a property of DDR5: its fit is set by a 128-core Xeon 6980P running llama.cpp (55.3 ms/token on a 3B model) and SGLang (17.0 ms) with the CPU executing every layer (research 12 §6). A GPU spilling into this memory pays it only for the share of each token read from RAM (lib/perf/decode.ts refinement 3). System RAM.active_share 0.7924 — placement per_token_overhead_ms 35.67706871212236 ms — derived: 0.208 × 1.58 ms (GPU VRAM) + 0.792 × 44.61 ms (System RAM) = 35.68 ms bytes_per_token_gb 18.4738 GB MoE: only active parameters are read per token (SPEC §3). Effective bits include embeddings and higher-precision tensors in mixed quants. Assumes active weights are quantized like the file average. KV cache is re-read on every token (DECISIONS D-008). Pools are read sequentially per token, so per-pool times add (harmonic combination). Per-token overhead on a split placement is DERIVED, not measured: each pool's own single-pool figure is charged in proportion to the share of each token it serves, because the fixed cost is per operation and the split puts that share of the operations on each device. Calibrated with that pool serving the whole model (D-013): GPU VRAM (1.58 ms), System RAM (44.61 ms). Batch size 1 (single user, interactive).	19.3 s (estimate) ESTIMATEQwen3-VL 235B-A22B Instruct time to first token on Hybrid · GeForce RTX 5090 32G GAMING TRIO OC + EPYC 9015 · 8× Server Premier KSM64R52BD4-96MC 96GB DDR5-6400 RDIMMArithmetic that also depends on a tunable assumption, listed below. ttft_s = prompt_tokens / prefill_tok_s = 32,768 / 1,699.82 = 19.3 s prompt_tokens 32768 tokens prefill_tok_s 1699.82 tok/s No published prefill measurement for this hardware; scaled from peak compute (estimate). Model q_dim unknown — attention FLOPs omitted, so long-prompt TTFT is understated.	$36,463 (derived) DERIVEDHybrid · GeForce RTX 5090 32G GAMING TRIO OC + EPYC 9015 · 8× Server Premier KSM64R52BD4-96MC 96GB DDR5-6400 RDIMM observed priceArithmetic on observed inputs only. No tunable assumption enters it. observed_price = Σ (cheapest observed unit price × quantity) $598 + $839 + $3,627 × 8 + $5,100 + $70 + $180 + $660 = $36,463 AMD EPYC 9015 59844 ¢ — SabrePC · new · tier 2/sabrepc_ngstate ASRock Rack GENOAD8QM3-2T/BCM 83900 ¢ — Newegg · new · tier 3/newegg_realtime Kingston Server Premier KSM64R52BD4-96MC 96GB DDR5-6400 RDIMM 362697 ¢ — SabrePC · new · tier 2/sabrepc_ngstate MSI GeForce RTX 5090 32G GAMING TRIO OC 509999 ¢ — Newegg · new · tier 3/newegg_realtime ARCTIC Freezer 4U-SP5 6999 ¢ — ARCTIC · new · tier 1/arctic_jsonld Phanteks Enthoo Pro 2 Server Edition 17999 ¢ — Newegg · new · tier 3/newegg_realtime Seasonic PRIME PX-2200 65999 ¢ — Newegg · new · tier 3/newegg_realtime Listed price only: shipping and sales tax are NOT in this figure. One price per part — the cheapest monitored, non-quarantined observation.
Hybrid · GeForce RTX 5090 32G GAMING TRIO OC + EPYC 9015 · 8× Server Premier KSM64R52BD4-128MB 128GB DDR5-6400 RDIMM 1,004 GB (derived) DERIVEDHybrid · GeForce RTX 5090 32G GAMING TRIO OC + EPYC 9015 · 8× Server Premier KSM64R52BD4-128MB 128GB DDR5-6400 RDIMM usable memoryArithmetic on observed inputs only. No tunable assumption enters it. available_gb = Σ usable_gb over memory pools GPU VRAM 31.04 + System RAM 972.8 = 1,003.84 GB GPU VRAM.usable_gb 31.04 GB — 32 GB physical System RAM.usable_gb 972.8 GB — 1,024 GB physical Usable memory is below physical memory for at least one pool (OS/driver reservation; editable in Settings). usable 1,792 (observed) OBSERVEDHybrid · GeForce RTX 5090 32G GAMING TRIO OC + EPYC 9015 · 8× Server Premier KSM64R52BD4-128MB 128GB DDR5-6400 RDIMM peak memory bandwidthMeasured or published, passed through unchanged — not computed here. peak memory bandwidth — published at the source, passed through unchanged; nothing is computed here 1,792 GB/s, as recorded from the GPU VRAM spec of Hybrid · GeForce RTX 5090 32G GAMING TRIO OC + EPYC 9015 · 8× Server Premier KSM64R52BD4-128MB 128GB DDR5-6400 RDIMM peak_memory_bandwidth 1,792 GB/s — the GPU VRAM spec of Hybrid · GeForce RTX 5090 32G GAMING TRIO OC + EPYC 9015 · 8× Server Premier KSM64R52BD4-128MB 128GB DDR5-6400 RDIMM Theoretical peak. Decode reaches a fraction of it — the efficiency factor is in the decode speed's own workings. The fastest pool. A machine with more than one pool places weights fastest-pool-first. Observed, not derived: this app records the figure and its source, and does not re-measure it. GB/s	Usable	9.2 tok/s (estimate) ESTIMATEQwen3-VL 235B-A22B Instruct decode speed on Hybrid · GeForce RTX 5090 32G GAMING TRIO OC + EPYC 9015 · 8× Server Premier KSM64R52BD4-128MB 128GB DDR5-6400 RDIMMArithmetic that also depends on a tunable assumption, listed below. tok_s = 1 / (Σ_pool bytes_per_token × share_pool / (bandwidth_pool × efficiency_pool) + Σ_pool share_pool × overhead_pool) GPU VRAM: 3.835 GB ÷ (1,792 GB/s × 0.48) + System RAM: 14.639 GB ÷ (307.2 GB/s × 0.7) + 35.68 ms fixed per-token overhead (share-weighted) → 9.24 tok/s GPU VRAM.bandwidth_gbs 1792 GB/s — component spec GPU VRAM.efficiency_moe 0.48 — Settings efficiency.gpu_vram — Discrete GPU: 10 dense + 3 MoE points; fit on published benchmarks, leave-one-out validated — docs/research/12b-estimator-validation.md GPU VRAM.active_share 0.2076 — placement System RAM.bandwidth_gbs 307.2 GB/s — component spec System RAM.efficiency_moe 0.7 — Settings efficiency.server_ddr5 — Server DDR5/MRDIMM single socket: 6 dense + 7 MoE points (small n); fit on published benchmarks, leave-one-out validated — docs/research/12b-estimator-validation.md. The 44.61 ms per-token overhead is a CPU-engine constant, not a property of DDR5: its fit is set by a 128-core Xeon 6980P running llama.cpp (55.3 ms/token on a 3B model) and SGLang (17.0 ms) with the CPU executing every layer (research 12 §6). A GPU spilling into this memory pays it only for the share of each token read from RAM (lib/perf/decode.ts refinement 3). System RAM.active_share 0.7924 — placement per_token_overhead_ms 35.67706871212236 ms — derived: 0.208 × 1.58 ms (GPU VRAM) + 0.792 × 44.61 ms (System RAM) = 35.68 ms bytes_per_token_gb 18.4738 GB MoE: only active parameters are read per token (SPEC §3). Effective bits include embeddings and higher-precision tensors in mixed quants. Assumes active weights are quantized like the file average. KV cache is re-read on every token (DECISIONS D-008). Pools are read sequentially per token, so per-pool times add (harmonic combination). Per-token overhead on a split placement is DERIVED, not measured: each pool's own single-pool figure is charged in proportion to the share of each token it serves, because the fixed cost is per operation and the split puts that share of the operations on each device. Calibrated with that pool serving the whole model (D-013): GPU VRAM (1.58 ms), System RAM (44.61 ms). Batch size 1 (single user, interactive).	19.3 s (estimate) ESTIMATEQwen3-VL 235B-A22B Instruct time to first token on Hybrid · GeForce RTX 5090 32G GAMING TRIO OC + EPYC 9015 · 8× Server Premier KSM64R52BD4-128MB 128GB DDR5-6400 RDIMMArithmetic that also depends on a tunable assumption, listed below. ttft_s = prompt_tokens / prefill_tok_s = 32,768 / 1,699.82 = 19.3 s prompt_tokens 32768 tokens prefill_tok_s 1699.82 tok/s No published prefill measurement for this hardware; scaled from peak compute (estimate). Model q_dim unknown — attention FLOPs omitted, so long-prompt TTFT is understated.	$38,872 (derived) DERIVEDHybrid · GeForce RTX 5090 32G GAMING TRIO OC + EPYC 9015 · 8× Server Premier KSM64R52BD4-128MB 128GB DDR5-6400 RDIMM observed priceArithmetic on observed inputs only. No tunable assumption enters it. observed_price = Σ (cheapest observed unit price × quantity) $598 + $839 + $3,928 × 8 + $5,100 + $70 + $180 + $660 = $38,872 AMD EPYC 9015 59844 ¢ — SabrePC · new · tier 2/sabrepc_ngstate ASRock Rack GENOAD8QM3-2T/BCM 83900 ¢ — Newegg · new · tier 3/newegg_realtime Kingston Server Premier KSM64R52BD4-128MB 128GB DDR5-6400 RDIMM 392811 ¢ — SabrePC · new · tier 2/sabrepc_ngstate MSI GeForce RTX 5090 32G GAMING TRIO OC 509999 ¢ — Newegg · new · tier 3/newegg_realtime ARCTIC Freezer 4U-SP5 6999 ¢ — ARCTIC · new · tier 1/arctic_jsonld Phanteks Enthoo Pro 2 Server Edition 17999 ¢ — Newegg · new · tier 3/newegg_realtime Seasonic PRIME PX-2200 65999 ¢ — Newegg · new · tier 3/newegg_realtime Listed price only: shipping and sales tax are NOT in this figure. One price per part — the cheapest monitored, non-quarantined observation.
2× GeForce RTX 5080 WINDFORCE OC 16G (GV-N5080WF3OC-16GD) · EPYC 9015 · 8× M321R4GA3BB6-CQK 32GB DDR5-4800 RDIMM 274 GB (derived) DERIVED2× GeForce RTX 5080 WINDFORCE OC 16G (GV-N5080WF3OC-16GD) · EPYC 9015 · 8× M321R4GA3BB6-CQK 32GB DDR5-4800 RDIMM usable memoryArithmetic on observed inputs only. No tunable assumption enters it. available_gb = Σ usable_gb over memory pools GPU VRAM 31.04 + System RAM 243.2 = 274.24 GB GPU VRAM.usable_gb 31.04 GB — 32 GB physical System RAM.usable_gb 243.2 GB — 256 GB physical Usable memory is below physical memory for at least one pool (OS/driver reservation; editable in Settings). usable 960 (observed) OBSERVED2× GeForce RTX 5080 WINDFORCE OC 16G (GV-N5080WF3OC-16GD) · EPYC 9015 · 8× M321R4GA3BB6-CQK 32GB DDR5-4800 RDIMM peak memory bandwidthMeasured or published, passed through unchanged — not computed here. peak memory bandwidth — published at the source, passed through unchanged; nothing is computed here 960 GB/s, as recorded from the GPU VRAM spec of 2× GeForce RTX 5080 WINDFORCE OC 16G (GV-N5080WF3OC-16GD) · EPYC 9015 · 8× M321R4GA3BB6-CQK 32GB DDR5-4800 RDIMM peak_memory_bandwidth 960 GB/s — the GPU VRAM spec of 2× GeForce RTX 5080 WINDFORCE OC 16G (GV-N5080WF3OC-16GD) · EPYC 9015 · 8× M321R4GA3BB6-CQK 32GB DDR5-4800 RDIMM Theoretical peak. Decode reaches a fraction of it — the efficiency factor is in the decode speed's own workings. The fastest pool. A machine with more than one pool places weights fastest-pool-first. Observed, not derived: this app records the figure and its source, and does not re-measure it. GB/s	Usable	8.9 tok/s (estimate) ESTIMATEQwen3-VL 235B-A22B Instruct decode speed on 2× GeForce RTX 5080 WINDFORCE OC 16G (GV-N5080WF3OC-16GD) · EPYC 9015 · 8× M321R4GA3BB6-CQK 32GB DDR5-4800 RDIMMArithmetic that also depends on a tunable assumption, listed below. tok_s = 1 / (Σ_pool bytes_per_token × share_pool / (bandwidth_pool × efficiency_pool) + Σ_pool share_pool × overhead_pool) GPU VRAM: 3.835 GB ÷ (960 GB/s × 0.48) + System RAM: 14.639 GB ÷ (307.2 GB/s × 0.7) + 35.68 ms fixed per-token overhead (share-weighted) → 8.92 tok/s GPU VRAM.bandwidth_gbs 960 GB/s — component spec GPU VRAM.efficiency_moe 0.48 — Settings efficiency.gpu_vram — Discrete GPU: 10 dense + 3 MoE points; fit on published benchmarks, leave-one-out validated — docs/research/12b-estimator-validation.md GPU VRAM.active_share 0.2076 — placement System RAM.bandwidth_gbs 307.2 GB/s — component spec System RAM.efficiency_moe 0.7 — Settings efficiency.server_ddr5 — Server DDR5/MRDIMM single socket: 6 dense + 7 MoE points (small n); fit on published benchmarks, leave-one-out validated — docs/research/12b-estimator-validation.md. The 44.61 ms per-token overhead is a CPU-engine constant, not a property of DDR5: its fit is set by a 128-core Xeon 6980P running llama.cpp (55.3 ms/token on a 3B model) and SGLang (17.0 ms) with the CPU executing every layer (research 12 §6). A GPU spilling into this memory pays it only for the share of each token read from RAM (lib/perf/decode.ts refinement 3). System RAM.active_share 0.7924 — placement per_token_overhead_ms 35.67706871212236 ms — derived: 0.208 × 1.58 ms (GPU VRAM) + 0.792 × 44.61 ms (System RAM) = 35.68 ms bytes_per_token_gb 18.4738 GB MoE: only active parameters are read per token (SPEC §3). Effective bits include embeddings and higher-precision tensors in mixed quants. Assumes active weights are quantized like the file average. KV cache is re-read on every token (DECISIONS D-008). Pools are read sequentially per token, so per-pool times add (harmonic combination). layer split across 2 GPUs: capacity adds, per-token bandwidth does not (layers run sequentially) Per-token overhead on a split placement is DERIVED, not measured: each pool's own single-pool figure is charged in proportion to the share of each token it serves, because the fixed cost is per operation and the split puts that share of the operations on each device. Calibrated with that pool serving the whole model (D-013): GPU VRAM (1.58 ms), System RAM (44.61 ms). Batch size 1 (single user, interactive).	35.9 s (estimate) ESTIMATEQwen3-VL 235B-A22B Instruct time to first token on 2× GeForce RTX 5080 WINDFORCE OC 16G (GV-N5080WF3OC-16GD) · EPYC 9015 · 8× M321R4GA3BB6-CQK 32GB DDR5-4800 RDIMMArithmetic that also depends on a tunable assumption, listed below. ttft_s = prompt_tokens / prefill_tok_s = 32,768 / 913.6 = 35.9 s prompt_tokens 32768 tokens prefill_tok_s 913.6 tok/s No published prefill measurement for this hardware; scaled from peak compute (estimate). Model q_dim unknown — attention FLOPs omitted, so long-prompt TTFT is understated.	$7,462 (derived) DERIVED2× GeForce RTX 5080 WINDFORCE OC 16G (GV-N5080WF3OC-16GD) · EPYC 9015 · 8× M321R4GA3BB6-CQK 32GB DDR5-4800 RDIMM observed priceArithmetic on observed inputs only. No tunable assumption enters it. observed_price = Σ (cheapest observed unit price × quantity) $598 + $839 + $214 × 8 + $1,700 × 2 + $70 + $180 + $660 = $7,462 AMD EPYC 9015 59844 ¢ — SabrePC · new · tier 2/sabrepc_ngstate ASRock Rack GENOAD8QM3-2T/BCM 83900 ¢ — Newegg · new · tier 3/newegg_realtime Samsung M321R4GA3BB6-CQK 32GB DDR5-4800 RDIMM 21427 ¢ — SabrePC · new · tier 2/sabrepc_ngstate GIGABYTE GeForce RTX 5080 WINDFORCE OC 16G (GV-N5080WF3OC-16GD) 169999 ¢ — Newegg · new · tier 3/newegg_realtime ARCTIC Freezer 4U-SP5 6999 ¢ — ARCTIC · new · tier 1/arctic_jsonld Phanteks Enthoo Pro 2 Server Edition 17999 ¢ — Newegg · new · tier 3/newegg_realtime Seasonic PRIME PX-2200 65999 ¢ — Newegg · new · tier 3/newegg_realtime Listed price only: shipping and sales tax are NOT in this figure. One price per part — the cheapest monitored, non-quarantined observation.
2× GeForce RTX 5080 16G SHADOW 3X OC · EPYC 9015 · 8× M321R4GA3BB6-CQK 32GB DDR5-4800 RDIMM 274 GB (derived) DERIVED2× GeForce RTX 5080 16G SHADOW 3X OC · EPYC 9015 · 8× M321R4GA3BB6-CQK 32GB DDR5-4800 RDIMM usable memoryArithmetic on observed inputs only. No tunable assumption enters it. available_gb = Σ usable_gb over memory pools GPU VRAM 31.04 + System RAM 243.2 = 274.24 GB GPU VRAM.usable_gb 31.04 GB — 32 GB physical System RAM.usable_gb 243.2 GB — 256 GB physical Usable memory is below physical memory for at least one pool (OS/driver reservation; editable in Settings). usable 960 (observed) OBSERVED2× GeForce RTX 5080 16G SHADOW 3X OC · EPYC 9015 · 8× M321R4GA3BB6-CQK 32GB DDR5-4800 RDIMM peak memory bandwidthMeasured or published, passed through unchanged — not computed here. peak memory bandwidth — published at the source, passed through unchanged; nothing is computed here 960 GB/s, as recorded from the GPU VRAM spec of 2× GeForce RTX 5080 16G SHADOW 3X OC · EPYC 9015 · 8× M321R4GA3BB6-CQK 32GB DDR5-4800 RDIMM peak_memory_bandwidth 960 GB/s — the GPU VRAM spec of 2× GeForce RTX 5080 16G SHADOW 3X OC · EPYC 9015 · 8× M321R4GA3BB6-CQK 32GB DDR5-4800 RDIMM Theoretical peak. Decode reaches a fraction of it — the efficiency factor is in the decode speed's own workings. The fastest pool. A machine with more than one pool places weights fastest-pool-first. Observed, not derived: this app records the figure and its source, and does not re-measure it. GB/s	Usable	8.9 tok/s (estimate) ESTIMATEQwen3-VL 235B-A22B Instruct decode speed on 2× GeForce RTX 5080 16G SHADOW 3X OC · EPYC 9015 · 8× M321R4GA3BB6-CQK 32GB DDR5-4800 RDIMMArithmetic that also depends on a tunable assumption, listed below. tok_s = 1 / (Σ_pool bytes_per_token × share_pool / (bandwidth_pool × efficiency_pool) + Σ_pool share_pool × overhead_pool) GPU VRAM: 3.835 GB ÷ (960 GB/s × 0.48) + System RAM: 14.639 GB ÷ (307.2 GB/s × 0.7) + 35.68 ms fixed per-token overhead (share-weighted) → 8.92 tok/s GPU VRAM.bandwidth_gbs 960 GB/s — component spec GPU VRAM.efficiency_moe 0.48 — Settings efficiency.gpu_vram — Discrete GPU: 10 dense + 3 MoE points; fit on published benchmarks, leave-one-out validated — docs/research/12b-estimator-validation.md GPU VRAM.active_share 0.2076 — placement System RAM.bandwidth_gbs 307.2 GB/s — component spec System RAM.efficiency_moe 0.7 — Settings efficiency.server_ddr5 — Server DDR5/MRDIMM single socket: 6 dense + 7 MoE points (small n); fit on published benchmarks, leave-one-out validated — docs/research/12b-estimator-validation.md. The 44.61 ms per-token overhead is a CPU-engine constant, not a property of DDR5: its fit is set by a 128-core Xeon 6980P running llama.cpp (55.3 ms/token on a 3B model) and SGLang (17.0 ms) with the CPU executing every layer (research 12 §6). A GPU spilling into this memory pays it only for the share of each token read from RAM (lib/perf/decode.ts refinement 3). System RAM.active_share 0.7924 — placement per_token_overhead_ms 35.67706871212236 ms — derived: 0.208 × 1.58 ms (GPU VRAM) + 0.792 × 44.61 ms (System RAM) = 35.68 ms bytes_per_token_gb 18.4738 GB MoE: only active parameters are read per token (SPEC §3). Effective bits include embeddings and higher-precision tensors in mixed quants. Assumes active weights are quantized like the file average. KV cache is re-read on every token (DECISIONS D-008). Pools are read sequentially per token, so per-pool times add (harmonic combination). layer split across 2 GPUs: capacity adds, per-token bandwidth does not (layers run sequentially) Per-token overhead on a split placement is DERIVED, not measured: each pool's own single-pool figure is charged in proportion to the share of each token it serves, because the fixed cost is per operation and the split puts that share of the operations on each device. Calibrated with that pool serving the whole model (D-013): GPU VRAM (1.58 ms), System RAM (44.61 ms). Batch size 1 (single user, interactive).	35.9 s (estimate) ESTIMATEQwen3-VL 235B-A22B Instruct time to first token on 2× GeForce RTX 5080 16G SHADOW 3X OC · EPYC 9015 · 8× M321R4GA3BB6-CQK 32GB DDR5-4800 RDIMMArithmetic that also depends on a tunable assumption, listed below. ttft_s = prompt_tokens / prefill_tok_s = 32,768 / 913.6 = 35.9 s prompt_tokens 32768 tokens prefill_tok_s 913.6 tok/s No published prefill measurement for this hardware; scaled from peak compute (estimate). Model q_dim unknown — attention FLOPs omitted, so long-prompt TTFT is understated.	$7,442 (derived) DERIVED2× GeForce RTX 5080 16G SHADOW 3X OC · EPYC 9015 · 8× M321R4GA3BB6-CQK 32GB DDR5-4800 RDIMM observed priceArithmetic on observed inputs only. No tunable assumption enters it. observed_price = Σ (cheapest observed unit price × quantity) $598 + $839 + $214 × 8 + $1,690 × 2 + $70 + $180 + $660 = $7,442 AMD EPYC 9015 59844 ¢ — SabrePC · new · tier 2/sabrepc_ngstate ASRock Rack GENOAD8QM3-2T/BCM 83900 ¢ — Newegg · new · tier 3/newegg_realtime Samsung M321R4GA3BB6-CQK 32GB DDR5-4800 RDIMM 21427 ¢ — SabrePC · new · tier 2/sabrepc_ngstate MSI GeForce RTX 5080 16G SHADOW 3X OC 168999 ¢ — Newegg · new · tier 3/newegg_realtime ARCTIC Freezer 4U-SP5 6999 ¢ — ARCTIC · new · tier 1/arctic_jsonld Phanteks Enthoo Pro 2 Server Edition 17999 ¢ — Newegg · new · tier 3/newegg_realtime Seasonic PRIME PX-2200 65999 ¢ — Newegg · new · tier 3/newegg_realtime Listed price only: shipping and sales tax are NOT in this figure. One price per part — the cheapest monitored, non-quarantined observation.
1× Radeon AI PRO R9700 Creator 32GB (R9700 CT) · EPYC 9015 · 8× M321R4GA3BB6-CQK 32GB DDR5-4800 RDIMM 274 GB (derived) DERIVED1× Radeon AI PRO R9700 Creator 32GB (R9700 CT) · EPYC 9015 · 8× M321R4GA3BB6-CQK 32GB DDR5-4800 RDIMM usable memoryArithmetic on observed inputs only. No tunable assumption enters it. available_gb = Σ usable_gb over memory pools GPU VRAM 31.04 + System RAM 243.2 = 274.24 GB GPU VRAM.usable_gb 31.04 GB — 32 GB physical System RAM.usable_gb 243.2 GB — 256 GB physical Usable memory is below physical memory for at least one pool (OS/driver reservation; editable in Settings). usable 640 (observed) OBSERVED1× Radeon AI PRO R9700 Creator 32GB (R9700 CT) · EPYC 9015 · 8× M321R4GA3BB6-CQK 32GB DDR5-4800 RDIMM peak memory bandwidthMeasured or published, passed through unchanged — not computed here. peak memory bandwidth — published at the source, passed through unchanged; nothing is computed here 640 GB/s, as recorded from the GPU VRAM spec of 1× Radeon AI PRO R9700 Creator 32GB (R9700 CT) · EPYC 9015 · 8× M321R4GA3BB6-CQK 32GB DDR5-4800 RDIMM peak_memory_bandwidth 640 GB/s — the GPU VRAM spec of 1× Radeon AI PRO R9700 Creator 32GB (R9700 CT) · EPYC 9015 · 8× M321R4GA3BB6-CQK 32GB DDR5-4800 RDIMM Theoretical peak. Decode reaches a fraction of it — the efficiency factor is in the decode speed's own workings. The fastest pool. A machine with more than one pool places weights fastest-pool-first. Observed, not derived: this app records the figure and its source, and does not re-measure it. GB/s	Usable	8.6 tok/s (estimate) ESTIMATEQwen3-VL 235B-A22B Instruct decode speed on 1× Radeon AI PRO R9700 Creator 32GB (R9700 CT) · EPYC 9015 · 8× M321R4GA3BB6-CQK 32GB DDR5-4800 RDIMMArithmetic that also depends on a tunable assumption, listed below. tok_s = 1 / (Σ_pool bytes_per_token × share_pool / (bandwidth_pool × efficiency_pool) + Σ_pool share_pool × overhead_pool) GPU VRAM: 3.835 GB ÷ (640 GB/s × 0.48) + System RAM: 14.639 GB ÷ (307.2 GB/s × 0.7) + 35.68 ms fixed per-token overhead (share-weighted) → 8.6 tok/s GPU VRAM.bandwidth_gbs 640 GB/s — component spec GPU VRAM.efficiency_moe 0.48 — Settings efficiency.gpu_vram — Discrete GPU: 10 dense + 3 MoE points; fit on published benchmarks, leave-one-out validated — docs/research/12b-estimator-validation.md GPU VRAM.active_share 0.2076 — placement System RAM.bandwidth_gbs 307.2 GB/s — component spec System RAM.efficiency_moe 0.7 — Settings efficiency.server_ddr5 — Server DDR5/MRDIMM single socket: 6 dense + 7 MoE points (small n); fit on published benchmarks, leave-one-out validated — docs/research/12b-estimator-validation.md. The 44.61 ms per-token overhead is a CPU-engine constant, not a property of DDR5: its fit is set by a 128-core Xeon 6980P running llama.cpp (55.3 ms/token on a 3B model) and SGLang (17.0 ms) with the CPU executing every layer (research 12 §6). A GPU spilling into this memory pays it only for the share of each token read from RAM (lib/perf/decode.ts refinement 3). System RAM.active_share 0.7924 — placement per_token_overhead_ms 35.67706871212236 ms — derived: 0.208 × 1.58 ms (GPU VRAM) + 0.792 × 44.61 ms (System RAM) = 35.68 ms bytes_per_token_gb 18.4738 GB MoE: only active parameters are read per token (SPEC §3). Effective bits include embeddings and higher-precision tensors in mixed quants. Assumes active weights are quantized like the file average. KV cache is re-read on every token (DECISIONS D-008). Pools are read sequentially per token, so per-pool times add (harmonic combination). Per-token overhead on a split placement is DERIVED, not measured: each pool's own single-pool figure is charged in proportion to the share of each token it serves, because the fixed cost is per operation and the split puts that share of the operations on each device. Calibrated with that pool serving the whole model (D-013): GPU VRAM (1.58 ms), System RAM (44.61 ms). Batch size 1 (single user, interactive).	21.1 s (estimate) ESTIMATEQwen3-VL 235B-A22B Instruct time to first token on 1× Radeon AI PRO R9700 Creator 32GB (R9700 CT) · EPYC 9015 · 8× M321R4GA3BB6-CQK 32GB DDR5-4800 RDIMMArithmetic that also depends on a tunable assumption, listed below. ttft_s = prompt_tokens / prefill_tok_s = 32,768 / 1,549.72 = 21.1 s prompt_tokens 32768 tokens prefill_tok_s 1549.72 tok/s No published prefill measurement for this hardware; scaled from peak compute (estimate). Model q_dim unknown — attention FLOPs omitted, so long-prompt TTFT is understated.	$5,762 (derived) DERIVED1× Radeon AI PRO R9700 Creator 32GB (R9700 CT) · EPYC 9015 · 8× M321R4GA3BB6-CQK 32GB DDR5-4800 RDIMM observed priceArithmetic on observed inputs only. No tunable assumption enters it. observed_price = Σ (cheapest observed unit price × quantity) $598 + $839 + $214 × 8 + $1,700 + $70 + $180 + $660 = $5,762 AMD EPYC 9015 59844 ¢ — SabrePC · new · tier 2/sabrepc_ngstate ASRock Rack GENOAD8QM3-2T/BCM 83900 ¢ — Newegg · new · tier 3/newegg_realtime Samsung M321R4GA3BB6-CQK 32GB DDR5-4800 RDIMM 21427 ¢ — SabrePC · new · tier 2/sabrepc_ngstate ASRock Radeon AI PRO R9700 Creator 32GB (R9700 CT) 169999 ¢ — Newegg · new · tier 2/seed:research ARCTIC Freezer 4U-SP5 6999 ¢ — ARCTIC · new · tier 1/arctic_jsonld Phanteks Enthoo Pro 2 Server Edition 17999 ¢ — Newegg · new · tier 3/newegg_realtime Seasonic PRIME PX-2200 65999 ¢ — Newegg · new · tier 3/newegg_realtime Listed price only: shipping and sales tax are NOT in this figure. One price per part — the cheapest monitored, non-quarantined observation.
1× Radeon AI PRO R9700 32GB (32358-01-20G) · EPYC 9015 · 8× M321R4GA3BB6-CQK 32GB DDR5-4800 RDIMM 274 GB (derived) DERIVED1× Radeon AI PRO R9700 32GB (32358-01-20G) · EPYC 9015 · 8× M321R4GA3BB6-CQK 32GB DDR5-4800 RDIMM usable memoryArithmetic on observed inputs only. No tunable assumption enters it. available_gb = Σ usable_gb over memory pools GPU VRAM 31.04 + System RAM 243.2 = 274.24 GB GPU VRAM.usable_gb 31.04 GB — 32 GB physical System RAM.usable_gb 243.2 GB — 256 GB physical Usable memory is below physical memory for at least one pool (OS/driver reservation; editable in Settings). usable 640 (observed) OBSERVED1× Radeon AI PRO R9700 32GB (32358-01-20G) · EPYC 9015 · 8× M321R4GA3BB6-CQK 32GB DDR5-4800 RDIMM peak memory bandwidthMeasured or published, passed through unchanged — not computed here. peak memory bandwidth — published at the source, passed through unchanged; nothing is computed here 640 GB/s, as recorded from the GPU VRAM spec of 1× Radeon AI PRO R9700 32GB (32358-01-20G) · EPYC 9015 · 8× M321R4GA3BB6-CQK 32GB DDR5-4800 RDIMM peak_memory_bandwidth 640 GB/s — the GPU VRAM spec of 1× Radeon AI PRO R9700 32GB (32358-01-20G) · EPYC 9015 · 8× M321R4GA3BB6-CQK 32GB DDR5-4800 RDIMM Theoretical peak. Decode reaches a fraction of it — the efficiency factor is in the decode speed's own workings. The fastest pool. A machine with more than one pool places weights fastest-pool-first. Observed, not derived: this app records the figure and its source, and does not re-measure it. GB/s	Usable	8.6 tok/s (estimate) ESTIMATEQwen3-VL 235B-A22B Instruct decode speed on 1× Radeon AI PRO R9700 32GB (32358-01-20G) · EPYC 9015 · 8× M321R4GA3BB6-CQK 32GB DDR5-4800 RDIMMArithmetic that also depends on a tunable assumption, listed below. tok_s = 1 / (Σ_pool bytes_per_token × share_pool / (bandwidth_pool × efficiency_pool) + Σ_pool share_pool × overhead_pool) GPU VRAM: 3.835 GB ÷ (640 GB/s × 0.48) + System RAM: 14.639 GB ÷ (307.2 GB/s × 0.7) + 35.68 ms fixed per-token overhead (share-weighted) → 8.6 tok/s GPU VRAM.bandwidth_gbs 640 GB/s — component spec GPU VRAM.efficiency_moe 0.48 — Settings efficiency.gpu_vram — Discrete GPU: 10 dense + 3 MoE points; fit on published benchmarks, leave-one-out validated — docs/research/12b-estimator-validation.md GPU VRAM.active_share 0.2076 — placement System RAM.bandwidth_gbs 307.2 GB/s — component spec System RAM.efficiency_moe 0.7 — Settings efficiency.server_ddr5 — Server DDR5/MRDIMM single socket: 6 dense + 7 MoE points (small n); fit on published benchmarks, leave-one-out validated — docs/research/12b-estimator-validation.md. The 44.61 ms per-token overhead is a CPU-engine constant, not a property of DDR5: its fit is set by a 128-core Xeon 6980P running llama.cpp (55.3 ms/token on a 3B model) and SGLang (17.0 ms) with the CPU executing every layer (research 12 §6). A GPU spilling into this memory pays it only for the share of each token read from RAM (lib/perf/decode.ts refinement 3). System RAM.active_share 0.7924 — placement per_token_overhead_ms 35.67706871212236 ms — derived: 0.208 × 1.58 ms (GPU VRAM) + 0.792 × 44.61 ms (System RAM) = 35.68 ms bytes_per_token_gb 18.4738 GB MoE: only active parameters are read per token (SPEC §3). Effective bits include embeddings and higher-precision tensors in mixed quants. Assumes active weights are quantized like the file average. KV cache is re-read on every token (DECISIONS D-008). Pools are read sequentially per token, so per-pool times add (harmonic combination). Per-token overhead on a split placement is DERIVED, not measured: each pool's own single-pool figure is charged in proportion to the share of each token it serves, because the fixed cost is per operation and the split puts that share of the operations on each device. Calibrated with that pool serving the whole model (D-013): GPU VRAM (1.58 ms), System RAM (44.61 ms). Batch size 1 (single user, interactive).	21.1 s (estimate) ESTIMATEQwen3-VL 235B-A22B Instruct time to first token on 1× Radeon AI PRO R9700 32GB (32358-01-20G) · EPYC 9015 · 8× M321R4GA3BB6-CQK 32GB DDR5-4800 RDIMMArithmetic that also depends on a tunable assumption, listed below. ttft_s = prompt_tokens / prefill_tok_s = 32,768 / 1,549.72 = 21.1 s prompt_tokens 32768 tokens prefill_tok_s 1549.72 tok/s No published prefill measurement for this hardware; scaled from peak compute (estimate). Model q_dim unknown — attention FLOPs omitted, so long-prompt TTFT is understated.	$5,666 (derived) DERIVED1× Radeon AI PRO R9700 32GB (32358-01-20G) · EPYC 9015 · 8× M321R4GA3BB6-CQK 32GB DDR5-4800 RDIMM observed priceArithmetic on observed inputs only. No tunable assumption enters it. observed_price = Σ (cheapest observed unit price × quantity) $598 + $839 + $214 × 8 + $1,604 + $70 + $180 + $660 = $5,666 AMD EPYC 9015 59844 ¢ — SabrePC · new · tier 2/sabrepc_ngstate ASRock Rack GENOAD8QM3-2T/BCM 83900 ¢ — Newegg · new · tier 3/newegg_realtime Samsung M321R4GA3BB6-CQK 32GB DDR5-4800 RDIMM 21427 ¢ — SabrePC · new · tier 2/sabrepc_ngstate Sapphire Radeon AI PRO R9700 32GB (32358-01-20G) 160431 ¢ — SabrePC · new · tier 2/sabrepc_ngstate ARCTIC Freezer 4U-SP5 6999 ¢ — ARCTIC · new · tier 1/arctic_jsonld Phanteks Enthoo Pro 2 Server Edition 17999 ¢ — Newegg · new · tier 3/newegg_realtime Seasonic PRIME PX-2200 65999 ¢ — Newegg · new · tier 3/newegg_realtime Listed price only: shipping and sales tax are NOT in this figure. One price per part — the cheapest monitored, non-quarantined observation.
1× AI TOP Radeon AI PRO R9700 32GB (GV-R9700AI TOP-32GD) · EPYC 9015 · 8× M321R4GA3BB6-CQK 32GB DDR5-4800 RDIMM 274 GB (derived) DERIVED1× AI TOP Radeon AI PRO R9700 32GB (GV-R9700AI TOP-32GD) · EPYC 9015 · 8× M321R4GA3BB6-CQK 32GB DDR5-4800 RDIMM usable memoryArithmetic on observed inputs only. No tunable assumption enters it. available_gb = Σ usable_gb over memory pools GPU VRAM 31.04 + System RAM 243.2 = 274.24 GB GPU VRAM.usable_gb 31.04 GB — 32 GB physical System RAM.usable_gb 243.2 GB — 256 GB physical Usable memory is below physical memory for at least one pool (OS/driver reservation; editable in Settings). usable 640 (observed) OBSERVED1× AI TOP Radeon AI PRO R9700 32GB (GV-R9700AI TOP-32GD) · EPYC 9015 · 8× M321R4GA3BB6-CQK 32GB DDR5-4800 RDIMM peak memory bandwidthMeasured or published, passed through unchanged — not computed here. peak memory bandwidth — published at the source, passed through unchanged; nothing is computed here 640 GB/s, as recorded from the GPU VRAM spec of 1× AI TOP Radeon AI PRO R9700 32GB (GV-R9700AI TOP-32GD) · EPYC 9015 · 8× M321R4GA3BB6-CQK 32GB DDR5-4800 RDIMM peak_memory_bandwidth 640 GB/s — the GPU VRAM spec of 1× AI TOP Radeon AI PRO R9700 32GB (GV-R9700AI TOP-32GD) · EPYC 9015 · 8× M321R4GA3BB6-CQK 32GB DDR5-4800 RDIMM Theoretical peak. Decode reaches a fraction of it — the efficiency factor is in the decode speed's own workings. The fastest pool. A machine with more than one pool places weights fastest-pool-first. Observed, not derived: this app records the figure and its source, and does not re-measure it. GB/s	Usable	8.6 tok/s (estimate) ESTIMATEQwen3-VL 235B-A22B Instruct decode speed on 1× AI TOP Radeon AI PRO R9700 32GB (GV-R9700AI TOP-32GD) · EPYC 9015 · 8× M321R4GA3BB6-CQK 32GB DDR5-4800 RDIMMArithmetic that also depends on a tunable assumption, listed below. tok_s = 1 / (Σ_pool bytes_per_token × share_pool / (bandwidth_pool × efficiency_pool) + Σ_pool share_pool × overhead_pool) GPU VRAM: 3.835 GB ÷ (640 GB/s × 0.48) + System RAM: 14.639 GB ÷ (307.2 GB/s × 0.7) + 35.68 ms fixed per-token overhead (share-weighted) → 8.6 tok/s GPU VRAM.bandwidth_gbs 640 GB/s — component spec GPU VRAM.efficiency_moe 0.48 — Settings efficiency.gpu_vram — Discrete GPU: 10 dense + 3 MoE points; fit on published benchmarks, leave-one-out validated — docs/research/12b-estimator-validation.md GPU VRAM.active_share 0.2076 — placement System RAM.bandwidth_gbs 307.2 GB/s — component spec System RAM.efficiency_moe 0.7 — Settings efficiency.server_ddr5 — Server DDR5/MRDIMM single socket: 6 dense + 7 MoE points (small n); fit on published benchmarks, leave-one-out validated — docs/research/12b-estimator-validation.md. The 44.61 ms per-token overhead is a CPU-engine constant, not a property of DDR5: its fit is set by a 128-core Xeon 6980P running llama.cpp (55.3 ms/token on a 3B model) and SGLang (17.0 ms) with the CPU executing every layer (research 12 §6). A GPU spilling into this memory pays it only for the share of each token read from RAM (lib/perf/decode.ts refinement 3). System RAM.active_share 0.7924 — placement per_token_overhead_ms 35.67706871212236 ms — derived: 0.208 × 1.58 ms (GPU VRAM) + 0.792 × 44.61 ms (System RAM) = 35.68 ms bytes_per_token_gb 18.4738 GB MoE: only active parameters are read per token (SPEC §3). Effective bits include embeddings and higher-precision tensors in mixed quants. Assumes active weights are quantized like the file average. KV cache is re-read on every token (DECISIONS D-008). Pools are read sequentially per token, so per-pool times add (harmonic combination). Per-token overhead on a split placement is DERIVED, not measured: each pool's own single-pool figure is charged in proportion to the share of each token it serves, because the fixed cost is per operation and the split puts that share of the operations on each device. Calibrated with that pool serving the whole model (D-013): GPU VRAM (1.58 ms), System RAM (44.61 ms). Batch size 1 (single user, interactive).	21.1 s (estimate) ESTIMATEQwen3-VL 235B-A22B Instruct time to first token on 1× AI TOP Radeon AI PRO R9700 32GB (GV-R9700AI TOP-32GD) · EPYC 9015 · 8× M321R4GA3BB6-CQK 32GB DDR5-4800 RDIMMArithmetic that also depends on a tunable assumption, listed below. ttft_s = prompt_tokens / prefill_tok_s = 32,768 / 1,549.72 = 21.1 s prompt_tokens 32768 tokens prefill_tok_s 1549.72 tok/s No published prefill measurement for this hardware; scaled from peak compute (estimate). Model q_dim unknown — attention FLOPs omitted, so long-prompt TTFT is understated.	$5,772 (derived) DERIVED1× AI TOP Radeon AI PRO R9700 32GB (GV-R9700AI TOP-32GD) · EPYC 9015 · 8× M321R4GA3BB6-CQK 32GB DDR5-4800 RDIMM observed priceArithmetic on observed inputs only. No tunable assumption enters it. observed_price = Σ (cheapest observed unit price × quantity) $598 + $839 + $214 × 8 + $1,711 + $70 + $180 + $660 = $5,772 AMD EPYC 9015 59844 ¢ — SabrePC · new · tier 2/sabrepc_ngstate ASRock Rack GENOAD8QM3-2T/BCM 83900 ¢ — Newegg · new · tier 3/newegg_realtime Samsung M321R4GA3BB6-CQK 32GB DDR5-4800 RDIMM 21427 ¢ — SabrePC · new · tier 2/sabrepc_ngstate GIGABYTE AI TOP Radeon AI PRO R9700 32GB (GV-R9700AI TOP-32GD) 171089 ¢ — SabrePC · new · tier 2/sabrepc_ngstate ARCTIC Freezer 4U-SP5 6999 ¢ — ARCTIC · new · tier 1/arctic_jsonld Phanteks Enthoo Pro 2 Server Edition 17999 ¢ — Newegg · new · tier 3/newegg_realtime Seasonic PRIME PX-2200 65999 ¢ — Newegg · new · tier 3/newegg_realtime Listed price only: shipping and sales tax are NOT in this figure. One price per part — the cheapest monitored, non-quarantined observation.
Hybrid · Radeon AI PRO R9700 Creator 32GB (R9700 CT) + EPYC 9015 · 8× M321R8GA0BB0-CQK 64GB DDR5-4800 RDIMM 517 GB (derived) DERIVEDHybrid · Radeon AI PRO R9700 Creator 32GB (R9700 CT) + EPYC 9015 · 8× M321R8GA0BB0-CQK 64GB DDR5-4800 RDIMM usable memoryArithmetic on observed inputs only. No tunable assumption enters it. available_gb = Σ usable_gb over memory pools GPU VRAM 31.04 + System RAM 486.4 = 517.44 GB GPU VRAM.usable_gb 31.04 GB — 32 GB physical System RAM.usable_gb 486.4 GB — 512 GB physical Usable memory is below physical memory for at least one pool (OS/driver reservation; editable in Settings). usable 640 (observed) OBSERVEDHybrid · Radeon AI PRO R9700 Creator 32GB (R9700 CT) + EPYC 9015 · 8× M321R8GA0BB0-CQK 64GB DDR5-4800 RDIMM peak memory bandwidthMeasured or published, passed through unchanged — not computed here. peak memory bandwidth — published at the source, passed through unchanged; nothing is computed here 640 GB/s, as recorded from the GPU VRAM spec of Hybrid · Radeon AI PRO R9700 Creator 32GB (R9700 CT) + EPYC 9015 · 8× M321R8GA0BB0-CQK 64GB DDR5-4800 RDIMM peak_memory_bandwidth 640 GB/s — the GPU VRAM spec of Hybrid · Radeon AI PRO R9700 Creator 32GB (R9700 CT) + EPYC 9015 · 8× M321R8GA0BB0-CQK 64GB DDR5-4800 RDIMM Theoretical peak. Decode reaches a fraction of it — the efficiency factor is in the decode speed's own workings. The fastest pool. A machine with more than one pool places weights fastest-pool-first. Observed, not derived: this app records the figure and its source, and does not re-measure it. GB/s	Usable	8.6 tok/s (estimate) ESTIMATEQwen3-VL 235B-A22B Instruct decode speed on Hybrid · Radeon AI PRO R9700 Creator 32GB (R9700 CT) + EPYC 9015 · 8× M321R8GA0BB0-CQK 64GB DDR5-4800 RDIMMArithmetic that also depends on a tunable assumption, listed below. tok_s = 1 / (Σ_pool bytes_per_token × share_pool / (bandwidth_pool × efficiency_pool) + Σ_pool share_pool × overhead_pool) GPU VRAM: 3.835 GB ÷ (640 GB/s × 0.48) + System RAM: 14.639 GB ÷ (307.2 GB/s × 0.7) + 35.68 ms fixed per-token overhead (share-weighted) → 8.6 tok/s GPU VRAM.bandwidth_gbs 640 GB/s — component spec GPU VRAM.efficiency_moe 0.48 — Settings efficiency.gpu_vram — Discrete GPU: 10 dense + 3 MoE points; fit on published benchmarks, leave-one-out validated — docs/research/12b-estimator-validation.md GPU VRAM.active_share 0.2076 — placement System RAM.bandwidth_gbs 307.2 GB/s — component spec System RAM.efficiency_moe 0.7 — Settings efficiency.server_ddr5 — Server DDR5/MRDIMM single socket: 6 dense + 7 MoE points (small n); fit on published benchmarks, leave-one-out validated — docs/research/12b-estimator-validation.md. The 44.61 ms per-token overhead is a CPU-engine constant, not a property of DDR5: its fit is set by a 128-core Xeon 6980P running llama.cpp (55.3 ms/token on a 3B model) and SGLang (17.0 ms) with the CPU executing every layer (research 12 §6). A GPU spilling into this memory pays it only for the share of each token read from RAM (lib/perf/decode.ts refinement 3). System RAM.active_share 0.7924 — placement per_token_overhead_ms 35.67706871212236 ms — derived: 0.208 × 1.58 ms (GPU VRAM) + 0.792 × 44.61 ms (System RAM) = 35.68 ms bytes_per_token_gb 18.4738 GB MoE: only active parameters are read per token (SPEC §3). Effective bits include embeddings and higher-precision tensors in mixed quants. Assumes active weights are quantized like the file average. KV cache is re-read on every token (DECISIONS D-008). Pools are read sequentially per token, so per-pool times add (harmonic combination). Per-token overhead on a split placement is DERIVED, not measured: each pool's own single-pool figure is charged in proportion to the share of each token it serves, because the fixed cost is per operation and the split puts that share of the operations on each device. Calibrated with that pool serving the whole model (D-013): GPU VRAM (1.58 ms), System RAM (44.61 ms). Batch size 1 (single user, interactive).	21.1 s (estimate) ESTIMATEQwen3-VL 235B-A22B Instruct time to first token on Hybrid · Radeon AI PRO R9700 Creator 32GB (R9700 CT) + EPYC 9015 · 8× M321R8GA0BB0-CQK 64GB DDR5-4800 RDIMMArithmetic that also depends on a tunable assumption, listed below. ttft_s = prompt_tokens / prefill_tok_s = 32,768 / 1,549.72 = 21.1 s prompt_tokens 32768 tokens prefill_tok_s 1549.72 tok/s No published prefill measurement for this hardware; scaled from peak compute (estimate). Model q_dim unknown — attention FLOPs omitted, so long-prompt TTFT is understated.	$6,424 (derived) DERIVEDHybrid · Radeon AI PRO R9700 Creator 32GB (R9700 CT) + EPYC 9015 · 8× M321R8GA0BB0-CQK 64GB DDR5-4800 RDIMM observed priceArithmetic on observed inputs only. No tunable assumption enters it. observed_price = Σ (cheapest observed unit price × quantity) $598 + $839 + $297 × 8 + $1,700 + $70 + $180 + $660 = $6,424 AMD EPYC 9015 59844 ¢ — SabrePC · new · tier 2/sabrepc_ngstate ASRock Rack GENOAD8QM3-2T/BCM 83900 ¢ — Newegg · new · tier 3/newegg_realtime Samsung M321R8GA0BB0-CQK 64GB DDR5-4800 RDIMM 29707 ¢ — SabrePC · new · tier 2/sabrepc_ngstate ASRock Radeon AI PRO R9700 Creator 32GB (R9700 CT) 169999 ¢ — Newegg · new · tier 2/seed:research ARCTIC Freezer 4U-SP5 6999 ¢ — ARCTIC · new · tier 1/arctic_jsonld Phanteks Enthoo Pro 2 Server Edition 17999 ¢ — Newegg · new · tier 3/newegg_realtime Seasonic PRIME PX-2200 65999 ¢ — Newegg · new · tier 3/newegg_realtime Listed price only: shipping and sales tax are NOT in this figure. One price per part — the cheapest monitored, non-quarantined observation.
Hybrid · Radeon AI PRO R9700 32GB (32358-01-20G) + EPYC 9015 · 8× M321R8GA0BB0-CQK 64GB DDR5-4800 RDIMM 517 GB (derived) DERIVEDHybrid · Radeon AI PRO R9700 32GB (32358-01-20G) + EPYC 9015 · 8× M321R8GA0BB0-CQK 64GB DDR5-4800 RDIMM usable memoryArithmetic on observed inputs only. No tunable assumption enters it. available_gb = Σ usable_gb over memory pools GPU VRAM 31.04 + System RAM 486.4 = 517.44 GB GPU VRAM.usable_gb 31.04 GB — 32 GB physical System RAM.usable_gb 486.4 GB — 512 GB physical Usable memory is below physical memory for at least one pool (OS/driver reservation; editable in Settings). usable 640 (observed) OBSERVEDHybrid · Radeon AI PRO R9700 32GB (32358-01-20G) + EPYC 9015 · 8× M321R8GA0BB0-CQK 64GB DDR5-4800 RDIMM peak memory bandwidthMeasured or published, passed through unchanged — not computed here. peak memory bandwidth — published at the source, passed through unchanged; nothing is computed here 640 GB/s, as recorded from the GPU VRAM spec of Hybrid · Radeon AI PRO R9700 32GB (32358-01-20G) + EPYC 9015 · 8× M321R8GA0BB0-CQK 64GB DDR5-4800 RDIMM peak_memory_bandwidth 640 GB/s — the GPU VRAM spec of Hybrid · Radeon AI PRO R9700 32GB (32358-01-20G) + EPYC 9015 · 8× M321R8GA0BB0-CQK 64GB DDR5-4800 RDIMM Theoretical peak. Decode reaches a fraction of it — the efficiency factor is in the decode speed's own workings. The fastest pool. A machine with more than one pool places weights fastest-pool-first. Observed, not derived: this app records the figure and its source, and does not re-measure it. GB/s	Usable	8.6 tok/s (estimate) ESTIMATEQwen3-VL 235B-A22B Instruct decode speed on Hybrid · Radeon AI PRO R9700 32GB (32358-01-20G) + EPYC 9015 · 8× M321R8GA0BB0-CQK 64GB DDR5-4800 RDIMMArithmetic that also depends on a tunable assumption, listed below. tok_s = 1 / (Σ_pool bytes_per_token × share_pool / (bandwidth_pool × efficiency_pool) + Σ_pool share_pool × overhead_pool) GPU VRAM: 3.835 GB ÷ (640 GB/s × 0.48) + System RAM: 14.639 GB ÷ (307.2 GB/s × 0.7) + 35.68 ms fixed per-token overhead (share-weighted) → 8.6 tok/s GPU VRAM.bandwidth_gbs 640 GB/s — component spec GPU VRAM.efficiency_moe 0.48 — Settings efficiency.gpu_vram — Discrete GPU: 10 dense + 3 MoE points; fit on published benchmarks, leave-one-out validated — docs/research/12b-estimator-validation.md GPU VRAM.active_share 0.2076 — placement System RAM.bandwidth_gbs 307.2 GB/s — component spec System RAM.efficiency_moe 0.7 — Settings efficiency.server_ddr5 — Server DDR5/MRDIMM single socket: 6 dense + 7 MoE points (small n); fit on published benchmarks, leave-one-out validated — docs/research/12b-estimator-validation.md. The 44.61 ms per-token overhead is a CPU-engine constant, not a property of DDR5: its fit is set by a 128-core Xeon 6980P running llama.cpp (55.3 ms/token on a 3B model) and SGLang (17.0 ms) with the CPU executing every layer (research 12 §6). A GPU spilling into this memory pays it only for the share of each token read from RAM (lib/perf/decode.ts refinement 3). System RAM.active_share 0.7924 — placement per_token_overhead_ms 35.67706871212236 ms — derived: 0.208 × 1.58 ms (GPU VRAM) + 0.792 × 44.61 ms (System RAM) = 35.68 ms bytes_per_token_gb 18.4738 GB MoE: only active parameters are read per token (SPEC §3). Effective bits include embeddings and higher-precision tensors in mixed quants. Assumes active weights are quantized like the file average. KV cache is re-read on every token (DECISIONS D-008). Pools are read sequentially per token, so per-pool times add (harmonic combination). Per-token overhead on a split placement is DERIVED, not measured: each pool's own single-pool figure is charged in proportion to the share of each token it serves, because the fixed cost is per operation and the split puts that share of the operations on each device. Calibrated with that pool serving the whole model (D-013): GPU VRAM (1.58 ms), System RAM (44.61 ms). Batch size 1 (single user, interactive).	21.1 s (estimate) ESTIMATEQwen3-VL 235B-A22B Instruct time to first token on Hybrid · Radeon AI PRO R9700 32GB (32358-01-20G) + EPYC 9015 · 8× M321R8GA0BB0-CQK 64GB DDR5-4800 RDIMMArithmetic that also depends on a tunable assumption, listed below. ttft_s = prompt_tokens / prefill_tok_s = 32,768 / 1,549.72 = 21.1 s prompt_tokens 32768 tokens prefill_tok_s 1549.72 tok/s No published prefill measurement for this hardware; scaled from peak compute (estimate). Model q_dim unknown — attention FLOPs omitted, so long-prompt TTFT is understated.	$6,328 (derived) DERIVEDHybrid · Radeon AI PRO R9700 32GB (32358-01-20G) + EPYC 9015 · 8× M321R8GA0BB0-CQK 64GB DDR5-4800 RDIMM observed priceArithmetic on observed inputs only. No tunable assumption enters it. observed_price = Σ (cheapest observed unit price × quantity) $598 + $839 + $297 × 8 + $1,604 + $70 + $180 + $660 = $6,328 AMD EPYC 9015 59844 ¢ — SabrePC · new · tier 2/sabrepc_ngstate ASRock Rack GENOAD8QM3-2T/BCM 83900 ¢ — Newegg · new · tier 3/newegg_realtime Samsung M321R8GA0BB0-CQK 64GB DDR5-4800 RDIMM 29707 ¢ — SabrePC · new · tier 2/sabrepc_ngstate Sapphire Radeon AI PRO R9700 32GB (32358-01-20G) 160431 ¢ — SabrePC · new · tier 2/sabrepc_ngstate ARCTIC Freezer 4U-SP5 6999 ¢ — ARCTIC · new · tier 1/arctic_jsonld Phanteks Enthoo Pro 2 Server Edition 17999 ¢ — Newegg · new · tier 3/newegg_realtime Seasonic PRIME PX-2200 65999 ¢ — Newegg · new · tier 3/newegg_realtime Listed price only: shipping and sales tax are NOT in this figure. One price per part — the cheapest monitored, non-quarantined observation.
Hybrid · AI TOP Radeon AI PRO R9700 32GB (GV-R9700AI TOP-32GD) + EPYC 9015 · 8× M321R8GA0BB0-CQK 64GB DDR5-4800 RDIMM 517 GB (derived) DERIVEDHybrid · AI TOP Radeon AI PRO R9700 32GB (GV-R9700AI TOP-32GD) + EPYC 9015 · 8× M321R8GA0BB0-CQK 64GB DDR5-4800 RDIMM usable memoryArithmetic on observed inputs only. No tunable assumption enters it. available_gb = Σ usable_gb over memory pools GPU VRAM 31.04 + System RAM 486.4 = 517.44 GB GPU VRAM.usable_gb 31.04 GB — 32 GB physical System RAM.usable_gb 486.4 GB — 512 GB physical Usable memory is below physical memory for at least one pool (OS/driver reservation; editable in Settings). usable 640 (observed) OBSERVEDHybrid · AI TOP Radeon AI PRO R9700 32GB (GV-R9700AI TOP-32GD) + EPYC 9015 · 8× M321R8GA0BB0-CQK 64GB DDR5-4800 RDIMM peak memory bandwidthMeasured or published, passed through unchanged — not computed here. peak memory bandwidth — published at the source, passed through unchanged; nothing is computed here 640 GB/s, as recorded from the GPU VRAM spec of Hybrid · AI TOP Radeon AI PRO R9700 32GB (GV-R9700AI TOP-32GD) + EPYC 9015 · 8× M321R8GA0BB0-CQK 64GB DDR5-4800 RDIMM peak_memory_bandwidth 640 GB/s — the GPU VRAM spec of Hybrid · AI TOP Radeon AI PRO R9700 32GB (GV-R9700AI TOP-32GD) + EPYC 9015 · 8× M321R8GA0BB0-CQK 64GB DDR5-4800 RDIMM Theoretical peak. Decode reaches a fraction of it — the efficiency factor is in the decode speed's own workings. The fastest pool. A machine with more than one pool places weights fastest-pool-first. Observed, not derived: this app records the figure and its source, and does not re-measure it. GB/s	Usable	8.6 tok/s (estimate) ESTIMATEQwen3-VL 235B-A22B Instruct decode speed on Hybrid · AI TOP Radeon AI PRO R9700 32GB (GV-R9700AI TOP-32GD) + EPYC 9015 · 8× M321R8GA0BB0-CQK 64GB DDR5-4800 RDIMMArithmetic that also depends on a tunable assumption, listed below. tok_s = 1 / (Σ_pool bytes_per_token × share_pool / (bandwidth_pool × efficiency_pool) + Σ_pool share_pool × overhead_pool) GPU VRAM: 3.835 GB ÷ (640 GB/s × 0.48) + System RAM: 14.639 GB ÷ (307.2 GB/s × 0.7) + 35.68 ms fixed per-token overhead (share-weighted) → 8.6 tok/s GPU VRAM.bandwidth_gbs 640 GB/s — component spec GPU VRAM.efficiency_moe 0.48 — Settings efficiency.gpu_vram — Discrete GPU: 10 dense + 3 MoE points; fit on published benchmarks, leave-one-out validated — docs/research/12b-estimator-validation.md GPU VRAM.active_share 0.2076 — placement System RAM.bandwidth_gbs 307.2 GB/s — component spec System RAM.efficiency_moe 0.7 — Settings efficiency.server_ddr5 — Server DDR5/MRDIMM single socket: 6 dense + 7 MoE points (small n); fit on published benchmarks, leave-one-out validated — docs/research/12b-estimator-validation.md. The 44.61 ms per-token overhead is a CPU-engine constant, not a property of DDR5: its fit is set by a 128-core Xeon 6980P running llama.cpp (55.3 ms/token on a 3B model) and SGLang (17.0 ms) with the CPU executing every layer (research 12 §6). A GPU spilling into this memory pays it only for the share of each token read from RAM (lib/perf/decode.ts refinement 3). System RAM.active_share 0.7924 — placement per_token_overhead_ms 35.67706871212236 ms — derived: 0.208 × 1.58 ms (GPU VRAM) + 0.792 × 44.61 ms (System RAM) = 35.68 ms bytes_per_token_gb 18.4738 GB MoE: only active parameters are read per token (SPEC §3). Effective bits include embeddings and higher-precision tensors in mixed quants. Assumes active weights are quantized like the file average. KV cache is re-read on every token (DECISIONS D-008). Pools are read sequentially per token, so per-pool times add (harmonic combination). Per-token overhead on a split placement is DERIVED, not measured: each pool's own single-pool figure is charged in proportion to the share of each token it serves, because the fixed cost is per operation and the split puts that share of the operations on each device. Calibrated with that pool serving the whole model (D-013): GPU VRAM (1.58 ms), System RAM (44.61 ms). Batch size 1 (single user, interactive).	21.1 s (estimate) ESTIMATEQwen3-VL 235B-A22B Instruct time to first token on Hybrid · AI TOP Radeon AI PRO R9700 32GB (GV-R9700AI TOP-32GD) + EPYC 9015 · 8× M321R8GA0BB0-CQK 64GB DDR5-4800 RDIMMArithmetic that also depends on a tunable assumption, listed below. ttft_s = prompt_tokens / prefill_tok_s = 32,768 / 1,549.72 = 21.1 s prompt_tokens 32768 tokens prefill_tok_s 1549.72 tok/s No published prefill measurement for this hardware; scaled from peak compute (estimate). Model q_dim unknown — attention FLOPs omitted, so long-prompt TTFT is understated.	$6,435 (derived) DERIVEDHybrid · AI TOP Radeon AI PRO R9700 32GB (GV-R9700AI TOP-32GD) + EPYC 9015 · 8× M321R8GA0BB0-CQK 64GB DDR5-4800 RDIMM observed priceArithmetic on observed inputs only. No tunable assumption enters it. observed_price = Σ (cheapest observed unit price × quantity) $598 + $839 + $297 × 8 + $1,711 + $70 + $180 + $660 = $6,435 AMD EPYC 9015 59844 ¢ — SabrePC · new · tier 2/sabrepc_ngstate ASRock Rack GENOAD8QM3-2T/BCM 83900 ¢ — Newegg · new · tier 3/newegg_realtime Samsung M321R8GA0BB0-CQK 64GB DDR5-4800 RDIMM 29707 ¢ — SabrePC · new · tier 2/sabrepc_ngstate GIGABYTE AI TOP Radeon AI PRO R9700 32GB (GV-R9700AI TOP-32GD) 171089 ¢ — SabrePC · new · tier 2/sabrepc_ngstate ARCTIC Freezer 4U-SP5 6999 ¢ — ARCTIC · new · tier 1/arctic_jsonld Phanteks Enthoo Pro 2 Server Edition 17999 ¢ — Newegg · new · tier 3/newegg_realtime Seasonic PRIME PX-2200 65999 ¢ — Newegg · new · tier 3/newegg_realtime Listed price only: shipping and sales tax are NOT in this figure. One price per part — the cheapest monitored, non-quarantined observation.
Hybrid · Radeon AI PRO R9700 Creator 32GB (R9700 CT) + EPYC 9015 · 8× Server Premier KSM64R52BD4-96MC 96GB DDR5-6400 RDIMM 761 GB (derived) DERIVEDHybrid · Radeon AI PRO R9700 Creator 32GB (R9700 CT) + EPYC 9015 · 8× Server Premier KSM64R52BD4-96MC 96GB DDR5-6400 RDIMM usable memoryArithmetic on observed inputs only. No tunable assumption enters it. available_gb = Σ usable_gb over memory pools GPU VRAM 31.04 + System RAM 729.6 = 760.64 GB GPU VRAM.usable_gb 31.04 GB — 32 GB physical System RAM.usable_gb 729.5999999999999 GB — 768 GB physical Usable memory is below physical memory for at least one pool (OS/driver reservation; editable in Settings). usable 640 (observed) OBSERVEDHybrid · Radeon AI PRO R9700 Creator 32GB (R9700 CT) + EPYC 9015 · 8× Server Premier KSM64R52BD4-96MC 96GB DDR5-6400 RDIMM peak memory bandwidthMeasured or published, passed through unchanged — not computed here. peak memory bandwidth — published at the source, passed through unchanged; nothing is computed here 640 GB/s, as recorded from the GPU VRAM spec of Hybrid · Radeon AI PRO R9700 Creator 32GB (R9700 CT) + EPYC 9015 · 8× Server Premier KSM64R52BD4-96MC 96GB DDR5-6400 RDIMM peak_memory_bandwidth 640 GB/s — the GPU VRAM spec of Hybrid · Radeon AI PRO R9700 Creator 32GB (R9700 CT) + EPYC 9015 · 8× Server Premier KSM64R52BD4-96MC 96GB DDR5-6400 RDIMM Theoretical peak. Decode reaches a fraction of it — the efficiency factor is in the decode speed's own workings. The fastest pool. A machine with more than one pool places weights fastest-pool-first. Observed, not derived: this app records the figure and its source, and does not re-measure it. GB/s	Usable	8.6 tok/s (estimate) ESTIMATEQwen3-VL 235B-A22B Instruct decode speed on Hybrid · Radeon AI PRO R9700 Creator 32GB (R9700 CT) + EPYC 9015 · 8× Server Premier KSM64R52BD4-96MC 96GB DDR5-6400 RDIMMArithmetic that also depends on a tunable assumption, listed below. tok_s = 1 / (Σ_pool bytes_per_token × share_pool / (bandwidth_pool × efficiency_pool) + Σ_pool share_pool × overhead_pool) GPU VRAM: 3.835 GB ÷ (640 GB/s × 0.48) + System RAM: 14.639 GB ÷ (307.2 GB/s × 0.7) + 35.68 ms fixed per-token overhead (share-weighted) → 8.6 tok/s GPU VRAM.bandwidth_gbs 640 GB/s — component spec GPU VRAM.efficiency_moe 0.48 — Settings efficiency.gpu_vram — Discrete GPU: 10 dense + 3 MoE points; fit on published benchmarks, leave-one-out validated — docs/research/12b-estimator-validation.md GPU VRAM.active_share 0.2076 — placement System RAM.bandwidth_gbs 307.2 GB/s — component spec System RAM.efficiency_moe 0.7 — Settings efficiency.server_ddr5 — Server DDR5/MRDIMM single socket: 6 dense + 7 MoE points (small n); fit on published benchmarks, leave-one-out validated — docs/research/12b-estimator-validation.md. The 44.61 ms per-token overhead is a CPU-engine constant, not a property of DDR5: its fit is set by a 128-core Xeon 6980P running llama.cpp (55.3 ms/token on a 3B model) and SGLang (17.0 ms) with the CPU executing every layer (research 12 §6). A GPU spilling into this memory pays it only for the share of each token read from RAM (lib/perf/decode.ts refinement 3). System RAM.active_share 0.7924 — placement per_token_overhead_ms 35.67706871212236 ms — derived: 0.208 × 1.58 ms (GPU VRAM) + 0.792 × 44.61 ms (System RAM) = 35.68 ms bytes_per_token_gb 18.4738 GB MoE: only active parameters are read per token (SPEC §3). Effective bits include embeddings and higher-precision tensors in mixed quants. Assumes active weights are quantized like the file average. KV cache is re-read on every token (DECISIONS D-008). Pools are read sequentially per token, so per-pool times add (harmonic combination). Per-token overhead on a split placement is DERIVED, not measured: each pool's own single-pool figure is charged in proportion to the share of each token it serves, because the fixed cost is per operation and the split puts that share of the operations on each device. Calibrated with that pool serving the whole model (D-013): GPU VRAM (1.58 ms), System RAM (44.61 ms). Batch size 1 (single user, interactive).	21.1 s (estimate) ESTIMATEQwen3-VL 235B-A22B Instruct time to first token on Hybrid · Radeon AI PRO R9700 Creator 32GB (R9700 CT) + EPYC 9015 · 8× Server Premier KSM64R52BD4-96MC 96GB DDR5-6400 RDIMMArithmetic that also depends on a tunable assumption, listed below. ttft_s = prompt_tokens / prefill_tok_s = 32,768 / 1,549.72 = 21.1 s prompt_tokens 32768 tokens prefill_tok_s 1549.72 tok/s No published prefill measurement for this hardware; scaled from peak compute (estimate). Model q_dim unknown — attention FLOPs omitted, so long-prompt TTFT is understated.	$33,063 (derived) DERIVEDHybrid · Radeon AI PRO R9700 Creator 32GB (R9700 CT) + EPYC 9015 · 8× Server Premier KSM64R52BD4-96MC 96GB DDR5-6400 RDIMM observed priceArithmetic on observed inputs only. No tunable assumption enters it. observed_price = Σ (cheapest observed unit price × quantity) $598 + $839 + $3,627 × 8 + $1,700 + $70 + $180 + $660 = $33,063 AMD EPYC 9015 59844 ¢ — SabrePC · new · tier 2/sabrepc_ngstate ASRock Rack GENOAD8QM3-2T/BCM 83900 ¢ — Newegg · new · tier 3/newegg_realtime Kingston Server Premier KSM64R52BD4-96MC 96GB DDR5-6400 RDIMM 362697 ¢ — SabrePC · new · tier 2/sabrepc_ngstate ASRock Radeon AI PRO R9700 Creator 32GB (R9700 CT) 169999 ¢ — Newegg · new · tier 2/seed:research ARCTIC Freezer 4U-SP5 6999 ¢ — ARCTIC · new · tier 1/arctic_jsonld Phanteks Enthoo Pro 2 Server Edition 17999 ¢ — Newegg · new · tier 3/newegg_realtime Seasonic PRIME PX-2200 65999 ¢ — Newegg · new · tier 3/newegg_realtime Listed price only: shipping and sales tax are NOT in this figure. One price per part — the cheapest monitored, non-quarantined observation.
Hybrid · Radeon AI PRO R9700 32GB (32358-01-20G) + EPYC 9015 · 8× Server Premier KSM64R52BD4-96MC 96GB DDR5-6400 RDIMM 761 GB (derived) DERIVEDHybrid · Radeon AI PRO R9700 32GB (32358-01-20G) + EPYC 9015 · 8× Server Premier KSM64R52BD4-96MC 96GB DDR5-6400 RDIMM usable memoryArithmetic on observed inputs only. No tunable assumption enters it. available_gb = Σ usable_gb over memory pools GPU VRAM 31.04 + System RAM 729.6 = 760.64 GB GPU VRAM.usable_gb 31.04 GB — 32 GB physical System RAM.usable_gb 729.5999999999999 GB — 768 GB physical Usable memory is below physical memory for at least one pool (OS/driver reservation; editable in Settings). usable 640 (observed) OBSERVEDHybrid · Radeon AI PRO R9700 32GB (32358-01-20G) + EPYC 9015 · 8× Server Premier KSM64R52BD4-96MC 96GB DDR5-6400 RDIMM peak memory bandwidthMeasured or published, passed through unchanged — not computed here. peak memory bandwidth — published at the source, passed through unchanged; nothing is computed here 640 GB/s, as recorded from the GPU VRAM spec of Hybrid · Radeon AI PRO R9700 32GB (32358-01-20G) + EPYC 9015 · 8× Server Premier KSM64R52BD4-96MC 96GB DDR5-6400 RDIMM peak_memory_bandwidth 640 GB/s — the GPU VRAM spec of Hybrid · Radeon AI PRO R9700 32GB (32358-01-20G) + EPYC 9015 · 8× Server Premier KSM64R52BD4-96MC 96GB DDR5-6400 RDIMM Theoretical peak. Decode reaches a fraction of it — the efficiency factor is in the decode speed's own workings. The fastest pool. A machine with more than one pool places weights fastest-pool-first. Observed, not derived: this app records the figure and its source, and does not re-measure it. GB/s	Usable	8.6 tok/s (estimate) ESTIMATEQwen3-VL 235B-A22B Instruct decode speed on Hybrid · Radeon AI PRO R9700 32GB (32358-01-20G) + EPYC 9015 · 8× Server Premier KSM64R52BD4-96MC 96GB DDR5-6400 RDIMMArithmetic that also depends on a tunable assumption, listed below. tok_s = 1 / (Σ_pool bytes_per_token × share_pool / (bandwidth_pool × efficiency_pool) + Σ_pool share_pool × overhead_pool) GPU VRAM: 3.835 GB ÷ (640 GB/s × 0.48) + System RAM: 14.639 GB ÷ (307.2 GB/s × 0.7) + 35.68 ms fixed per-token overhead (share-weighted) → 8.6 tok/s GPU VRAM.bandwidth_gbs 640 GB/s — component spec GPU VRAM.efficiency_moe 0.48 — Settings efficiency.gpu_vram — Discrete GPU: 10 dense + 3 MoE points; fit on published benchmarks, leave-one-out validated — docs/research/12b-estimator-validation.md GPU VRAM.active_share 0.2076 — placement System RAM.bandwidth_gbs 307.2 GB/s — component spec System RAM.efficiency_moe 0.7 — Settings efficiency.server_ddr5 — Server DDR5/MRDIMM single socket: 6 dense + 7 MoE points (small n); fit on published benchmarks, leave-one-out validated — docs/research/12b-estimator-validation.md. The 44.61 ms per-token overhead is a CPU-engine constant, not a property of DDR5: its fit is set by a 128-core Xeon 6980P running llama.cpp (55.3 ms/token on a 3B model) and SGLang (17.0 ms) with the CPU executing every layer (research 12 §6). A GPU spilling into this memory pays it only for the share of each token read from RAM (lib/perf/decode.ts refinement 3). System RAM.active_share 0.7924 — placement per_token_overhead_ms 35.67706871212236 ms — derived: 0.208 × 1.58 ms (GPU VRAM) + 0.792 × 44.61 ms (System RAM) = 35.68 ms bytes_per_token_gb 18.4738 GB MoE: only active parameters are read per token (SPEC §3). Effective bits include embeddings and higher-precision tensors in mixed quants. Assumes active weights are quantized like the file average. KV cache is re-read on every token (DECISIONS D-008). Pools are read sequentially per token, so per-pool times add (harmonic combination). Per-token overhead on a split placement is DERIVED, not measured: each pool's own single-pool figure is charged in proportion to the share of each token it serves, because the fixed cost is per operation and the split puts that share of the operations on each device. Calibrated with that pool serving the whole model (D-013): GPU VRAM (1.58 ms), System RAM (44.61 ms). Batch size 1 (single user, interactive).	21.1 s (estimate) ESTIMATEQwen3-VL 235B-A22B Instruct time to first token on Hybrid · Radeon AI PRO R9700 32GB (32358-01-20G) + EPYC 9015 · 8× Server Premier KSM64R52BD4-96MC 96GB DDR5-6400 RDIMMArithmetic that also depends on a tunable assumption, listed below. ttft_s = prompt_tokens / prefill_tok_s = 32,768 / 1,549.72 = 21.1 s prompt_tokens 32768 tokens prefill_tok_s 1549.72 tok/s No published prefill measurement for this hardware; scaled from peak compute (estimate). Model q_dim unknown — attention FLOPs omitted, so long-prompt TTFT is understated.	$32,967 (derived) DERIVEDHybrid · Radeon AI PRO R9700 32GB (32358-01-20G) + EPYC 9015 · 8× Server Premier KSM64R52BD4-96MC 96GB DDR5-6400 RDIMM observed priceArithmetic on observed inputs only. No tunable assumption enters it. observed_price = Σ (cheapest observed unit price × quantity) $598 + $839 + $3,627 × 8 + $1,604 + $70 + $180 + $660 = $32,967 AMD EPYC 9015 59844 ¢ — SabrePC · new · tier 2/sabrepc_ngstate ASRock Rack GENOAD8QM3-2T/BCM 83900 ¢ — Newegg · new · tier 3/newegg_realtime Kingston Server Premier KSM64R52BD4-96MC 96GB DDR5-6400 RDIMM 362697 ¢ — SabrePC · new · tier 2/sabrepc_ngstate Sapphire Radeon AI PRO R9700 32GB (32358-01-20G) 160431 ¢ — SabrePC · new · tier 2/sabrepc_ngstate ARCTIC Freezer 4U-SP5 6999 ¢ — ARCTIC · new · tier 1/arctic_jsonld Phanteks Enthoo Pro 2 Server Edition 17999 ¢ — Newegg · new · tier 3/newegg_realtime Seasonic PRIME PX-2200 65999 ¢ — Newegg · new · tier 3/newegg_realtime Listed price only: shipping and sales tax are NOT in this figure. One price per part — the cheapest monitored, non-quarantined observation.
Hybrid · AI TOP Radeon AI PRO R9700 32GB (GV-R9700AI TOP-32GD) + EPYC 9015 · 8× Server Premier KSM64R52BD4-96MC 96GB DDR5-6400 RDIMM 761 GB (derived) DERIVEDHybrid · AI TOP Radeon AI PRO R9700 32GB (GV-R9700AI TOP-32GD) + EPYC 9015 · 8× Server Premier KSM64R52BD4-96MC 96GB DDR5-6400 RDIMM usable memoryArithmetic on observed inputs only. No tunable assumption enters it. available_gb = Σ usable_gb over memory pools GPU VRAM 31.04 + System RAM 729.6 = 760.64 GB GPU VRAM.usable_gb 31.04 GB — 32 GB physical System RAM.usable_gb 729.5999999999999 GB — 768 GB physical Usable memory is below physical memory for at least one pool (OS/driver reservation; editable in Settings). usable 640 (observed) OBSERVEDHybrid · AI TOP Radeon AI PRO R9700 32GB (GV-R9700AI TOP-32GD) + EPYC 9015 · 8× Server Premier KSM64R52BD4-96MC 96GB DDR5-6400 RDIMM peak memory bandwidthMeasured or published, passed through unchanged — not computed here. peak memory bandwidth — published at the source, passed through unchanged; nothing is computed here 640 GB/s, as recorded from the GPU VRAM spec of Hybrid · AI TOP Radeon AI PRO R9700 32GB (GV-R9700AI TOP-32GD) + EPYC 9015 · 8× Server Premier KSM64R52BD4-96MC 96GB DDR5-6400 RDIMM peak_memory_bandwidth 640 GB/s — the GPU VRAM spec of Hybrid · AI TOP Radeon AI PRO R9700 32GB (GV-R9700AI TOP-32GD) + EPYC 9015 · 8× Server Premier KSM64R52BD4-96MC 96GB DDR5-6400 RDIMM Theoretical peak. Decode reaches a fraction of it — the efficiency factor is in the decode speed's own workings. The fastest pool. A machine with more than one pool places weights fastest-pool-first. Observed, not derived: this app records the figure and its source, and does not re-measure it. GB/s	Usable	8.6 tok/s (estimate) ESTIMATEQwen3-VL 235B-A22B Instruct decode speed on Hybrid · AI TOP Radeon AI PRO R9700 32GB (GV-R9700AI TOP-32GD) + EPYC 9015 · 8× Server Premier KSM64R52BD4-96MC 96GB DDR5-6400 RDIMMArithmetic that also depends on a tunable assumption, listed below. tok_s = 1 / (Σ_pool bytes_per_token × share_pool / (bandwidth_pool × efficiency_pool) + Σ_pool share_pool × overhead_pool) GPU VRAM: 3.835 GB ÷ (640 GB/s × 0.48) + System RAM: 14.639 GB ÷ (307.2 GB/s × 0.7) + 35.68 ms fixed per-token overhead (share-weighted) → 8.6 tok/s GPU VRAM.bandwidth_gbs 640 GB/s — component spec GPU VRAM.efficiency_moe 0.48 — Settings efficiency.gpu_vram — Discrete GPU: 10 dense + 3 MoE points; fit on published benchmarks, leave-one-out validated — docs/research/12b-estimator-validation.md GPU VRAM.active_share 0.2076 — placement System RAM.bandwidth_gbs 307.2 GB/s — component spec System RAM.efficiency_moe 0.7 — Settings efficiency.server_ddr5 — Server DDR5/MRDIMM single socket: 6 dense + 7 MoE points (small n); fit on published benchmarks, leave-one-out validated — docs/research/12b-estimator-validation.md. The 44.61 ms per-token overhead is a CPU-engine constant, not a property of DDR5: its fit is set by a 128-core Xeon 6980P running llama.cpp (55.3 ms/token on a 3B model) and SGLang (17.0 ms) with the CPU executing every layer (research 12 §6). A GPU spilling into this memory pays it only for the share of each token read from RAM (lib/perf/decode.ts refinement 3). System RAM.active_share 0.7924 — placement per_token_overhead_ms 35.67706871212236 ms — derived: 0.208 × 1.58 ms (GPU VRAM) + 0.792 × 44.61 ms (System RAM) = 35.68 ms bytes_per_token_gb 18.4738 GB MoE: only active parameters are read per token (SPEC §3). Effective bits include embeddings and higher-precision tensors in mixed quants. Assumes active weights are quantized like the file average. KV cache is re-read on every token (DECISIONS D-008). Pools are read sequentially per token, so per-pool times add (harmonic combination). Per-token overhead on a split placement is DERIVED, not measured: each pool's own single-pool figure is charged in proportion to the share of each token it serves, because the fixed cost is per operation and the split puts that share of the operations on each device. Calibrated with that pool serving the whole model (D-013): GPU VRAM (1.58 ms), System RAM (44.61 ms). Batch size 1 (single user, interactive).	21.1 s (estimate) ESTIMATEQwen3-VL 235B-A22B Instruct time to first token on Hybrid · AI TOP Radeon AI PRO R9700 32GB (GV-R9700AI TOP-32GD) + EPYC 9015 · 8× Server Premier KSM64R52BD4-96MC 96GB DDR5-6400 RDIMMArithmetic that also depends on a tunable assumption, listed below. ttft_s = prompt_tokens / prefill_tok_s = 32,768 / 1,549.72 = 21.1 s prompt_tokens 32768 tokens prefill_tok_s 1549.72 tok/s No published prefill measurement for this hardware; scaled from peak compute (estimate). Model q_dim unknown — attention FLOPs omitted, so long-prompt TTFT is understated.	$33,074 (derived) DERIVEDHybrid · AI TOP Radeon AI PRO R9700 32GB (GV-R9700AI TOP-32GD) + EPYC 9015 · 8× Server Premier KSM64R52BD4-96MC 96GB DDR5-6400 RDIMM observed priceArithmetic on observed inputs only. No tunable assumption enters it. observed_price = Σ (cheapest observed unit price × quantity) $598 + $839 + $3,627 × 8 + $1,711 + $70 + $180 + $660 = $33,074 AMD EPYC 9015 59844 ¢ — SabrePC · new · tier 2/sabrepc_ngstate ASRock Rack GENOAD8QM3-2T/BCM 83900 ¢ — Newegg · new · tier 3/newegg_realtime Kingston Server Premier KSM64R52BD4-96MC 96GB DDR5-6400 RDIMM 362697 ¢ — SabrePC · new · tier 2/sabrepc_ngstate GIGABYTE AI TOP Radeon AI PRO R9700 32GB (GV-R9700AI TOP-32GD) 171089 ¢ — SabrePC · new · tier 2/sabrepc_ngstate ARCTIC Freezer 4U-SP5 6999 ¢ — ARCTIC · new · tier 1/arctic_jsonld Phanteks Enthoo Pro 2 Server Edition 17999 ¢ — Newegg · new · tier 3/newegg_realtime Seasonic PRIME PX-2200 65999 ¢ — Newegg · new · tier 3/newegg_realtime Listed price only: shipping and sales tax are NOT in this figure. One price per part — the cheapest monitored, non-quarantined observation.
Hybrid · Radeon AI PRO R9700 Creator 32GB (R9700 CT) + EPYC 9015 · 8× Server Premier KSM64R52BD4-128MB 128GB DDR5-6400 RDIMM 1,004 GB (derived) DERIVEDHybrid · Radeon AI PRO R9700 Creator 32GB (R9700 CT) + EPYC 9015 · 8× Server Premier KSM64R52BD4-128MB 128GB DDR5-6400 RDIMM usable memoryArithmetic on observed inputs only. No tunable assumption enters it. available_gb = Σ usable_gb over memory pools GPU VRAM 31.04 + System RAM 972.8 = 1,003.84 GB GPU VRAM.usable_gb 31.04 GB — 32 GB physical System RAM.usable_gb 972.8 GB — 1,024 GB physical Usable memory is below physical memory for at least one pool (OS/driver reservation; editable in Settings). usable 640 (observed) OBSERVEDHybrid · Radeon AI PRO R9700 Creator 32GB (R9700 CT) + EPYC 9015 · 8× Server Premier KSM64R52BD4-128MB 128GB DDR5-6400 RDIMM peak memory bandwidthMeasured or published, passed through unchanged — not computed here. peak memory bandwidth — published at the source, passed through unchanged; nothing is computed here 640 GB/s, as recorded from the GPU VRAM spec of Hybrid · Radeon AI PRO R9700 Creator 32GB (R9700 CT) + EPYC 9015 · 8× Server Premier KSM64R52BD4-128MB 128GB DDR5-6400 RDIMM peak_memory_bandwidth 640 GB/s — the GPU VRAM spec of Hybrid · Radeon AI PRO R9700 Creator 32GB (R9700 CT) + EPYC 9015 · 8× Server Premier KSM64R52BD4-128MB 128GB DDR5-6400 RDIMM Theoretical peak. Decode reaches a fraction of it — the efficiency factor is in the decode speed's own workings. The fastest pool. A machine with more than one pool places weights fastest-pool-first. Observed, not derived: this app records the figure and its source, and does not re-measure it. GB/s	Usable	8.6 tok/s (estimate) ESTIMATEQwen3-VL 235B-A22B Instruct decode speed on Hybrid · Radeon AI PRO R9700 Creator 32GB (R9700 CT) + EPYC 9015 · 8× Server Premier KSM64R52BD4-128MB 128GB DDR5-6400 RDIMMArithmetic that also depends on a tunable assumption, listed below. tok_s = 1 / (Σ_pool bytes_per_token × share_pool / (bandwidth_pool × efficiency_pool) + Σ_pool share_pool × overhead_pool) GPU VRAM: 3.835 GB ÷ (640 GB/s × 0.48) + System RAM: 14.639 GB ÷ (307.2 GB/s × 0.7) + 35.68 ms fixed per-token overhead (share-weighted) → 8.6 tok/s GPU VRAM.bandwidth_gbs 640 GB/s — component spec GPU VRAM.efficiency_moe 0.48 — Settings efficiency.gpu_vram — Discrete GPU: 10 dense + 3 MoE points; fit on published benchmarks, leave-one-out validated — docs/research/12b-estimator-validation.md GPU VRAM.active_share 0.2076 — placement System RAM.bandwidth_gbs 307.2 GB/s — component spec System RAM.efficiency_moe 0.7 — Settings efficiency.server_ddr5 — Server DDR5/MRDIMM single socket: 6 dense + 7 MoE points (small n); fit on published benchmarks, leave-one-out validated — docs/research/12b-estimator-validation.md. The 44.61 ms per-token overhead is a CPU-engine constant, not a property of DDR5: its fit is set by a 128-core Xeon 6980P running llama.cpp (55.3 ms/token on a 3B model) and SGLang (17.0 ms) with the CPU executing every layer (research 12 §6). A GPU spilling into this memory pays it only for the share of each token read from RAM (lib/perf/decode.ts refinement 3). System RAM.active_share 0.7924 — placement per_token_overhead_ms 35.67706871212236 ms — derived: 0.208 × 1.58 ms (GPU VRAM) + 0.792 × 44.61 ms (System RAM) = 35.68 ms bytes_per_token_gb 18.4738 GB MoE: only active parameters are read per token (SPEC §3). Effective bits include embeddings and higher-precision tensors in mixed quants. Assumes active weights are quantized like the file average. KV cache is re-read on every token (DECISIONS D-008). Pools are read sequentially per token, so per-pool times add (harmonic combination). Per-token overhead on a split placement is DERIVED, not measured: each pool's own single-pool figure is charged in proportion to the share of each token it serves, because the fixed cost is per operation and the split puts that share of the operations on each device. Calibrated with that pool serving the whole model (D-013): GPU VRAM (1.58 ms), System RAM (44.61 ms). Batch size 1 (single user, interactive).	21.1 s (estimate) ESTIMATEQwen3-VL 235B-A22B Instruct time to first token on Hybrid · Radeon AI PRO R9700 Creator 32GB (R9700 CT) + EPYC 9015 · 8× Server Premier KSM64R52BD4-128MB 128GB DDR5-6400 RDIMMArithmetic that also depends on a tunable assumption, listed below. ttft_s = prompt_tokens / prefill_tok_s = 32,768 / 1,549.72 = 21.1 s prompt_tokens 32768 tokens prefill_tok_s 1549.72 tok/s No published prefill measurement for this hardware; scaled from peak compute (estimate). Model q_dim unknown — attention FLOPs omitted, so long-prompt TTFT is understated.	$35,472 (derived) DERIVEDHybrid · Radeon AI PRO R9700 Creator 32GB (R9700 CT) + EPYC 9015 · 8× Server Premier KSM64R52BD4-128MB 128GB DDR5-6400 RDIMM observed priceArithmetic on observed inputs only. No tunable assumption enters it. observed_price = Σ (cheapest observed unit price × quantity) $598 + $839 + $3,928 × 8 + $1,700 + $70 + $180 + $660 = $35,472 AMD EPYC 9015 59844 ¢ — SabrePC · new · tier 2/sabrepc_ngstate ASRock Rack GENOAD8QM3-2T/BCM 83900 ¢ — Newegg · new · tier 3/newegg_realtime Kingston Server Premier KSM64R52BD4-128MB 128GB DDR5-6400 RDIMM 392811 ¢ — SabrePC · new · tier 2/sabrepc_ngstate ASRock Radeon AI PRO R9700 Creator 32GB (R9700 CT) 169999 ¢ — Newegg · new · tier 2/seed:research ARCTIC Freezer 4U-SP5 6999 ¢ — ARCTIC · new · tier 1/arctic_jsonld Phanteks Enthoo Pro 2 Server Edition 17999 ¢ — Newegg · new · tier 3/newegg_realtime Seasonic PRIME PX-2200 65999 ¢ — Newegg · new · tier 3/newegg_realtime Listed price only: shipping and sales tax are NOT in this figure. One price per part — the cheapest monitored, non-quarantined observation.
Hybrid · Radeon AI PRO R9700 32GB (32358-01-20G) + EPYC 9015 · 8× Server Premier KSM64R52BD4-128MB 128GB DDR5-6400 RDIMM 1,004 GB (derived) DERIVEDHybrid · Radeon AI PRO R9700 32GB (32358-01-20G) + EPYC 9015 · 8× Server Premier KSM64R52BD4-128MB 128GB DDR5-6400 RDIMM usable memoryArithmetic on observed inputs only. No tunable assumption enters it. available_gb = Σ usable_gb over memory pools GPU VRAM 31.04 + System RAM 972.8 = 1,003.84 GB GPU VRAM.usable_gb 31.04 GB — 32 GB physical System RAM.usable_gb 972.8 GB — 1,024 GB physical Usable memory is below physical memory for at least one pool (OS/driver reservation; editable in Settings). usable 640 (observed) OBSERVEDHybrid · Radeon AI PRO R9700 32GB (32358-01-20G) + EPYC 9015 · 8× Server Premier KSM64R52BD4-128MB 128GB DDR5-6400 RDIMM peak memory bandwidthMeasured or published, passed through unchanged — not computed here. peak memory bandwidth — published at the source, passed through unchanged; nothing is computed here 640 GB/s, as recorded from the GPU VRAM spec of Hybrid · Radeon AI PRO R9700 32GB (32358-01-20G) + EPYC 9015 · 8× Server Premier KSM64R52BD4-128MB 128GB DDR5-6400 RDIMM peak_memory_bandwidth 640 GB/s — the GPU VRAM spec of Hybrid · Radeon AI PRO R9700 32GB (32358-01-20G) + EPYC 9015 · 8× Server Premier KSM64R52BD4-128MB 128GB DDR5-6400 RDIMM Theoretical peak. Decode reaches a fraction of it — the efficiency factor is in the decode speed's own workings. The fastest pool. A machine with more than one pool places weights fastest-pool-first. Observed, not derived: this app records the figure and its source, and does not re-measure it. GB/s	Usable	8.6 tok/s (estimate) ESTIMATEQwen3-VL 235B-A22B Instruct decode speed on Hybrid · Radeon AI PRO R9700 32GB (32358-01-20G) + EPYC 9015 · 8× Server Premier KSM64R52BD4-128MB 128GB DDR5-6400 RDIMMArithmetic that also depends on a tunable assumption, listed below. tok_s = 1 / (Σ_pool bytes_per_token × share_pool / (bandwidth_pool × efficiency_pool) + Σ_pool share_pool × overhead_pool) GPU VRAM: 3.835 GB ÷ (640 GB/s × 0.48) + System RAM: 14.639 GB ÷ (307.2 GB/s × 0.7) + 35.68 ms fixed per-token overhead (share-weighted) → 8.6 tok/s GPU VRAM.bandwidth_gbs 640 GB/s — component spec GPU VRAM.efficiency_moe 0.48 — Settings efficiency.gpu_vram — Discrete GPU: 10 dense + 3 MoE points; fit on published benchmarks, leave-one-out validated — docs/research/12b-estimator-validation.md GPU VRAM.active_share 0.2076 — placement System RAM.bandwidth_gbs 307.2 GB/s — component spec System RAM.efficiency_moe 0.7 — Settings efficiency.server_ddr5 — Server DDR5/MRDIMM single socket: 6 dense + 7 MoE points (small n); fit on published benchmarks, leave-one-out validated — docs/research/12b-estimator-validation.md. The 44.61 ms per-token overhead is a CPU-engine constant, not a property of DDR5: its fit is set by a 128-core Xeon 6980P running llama.cpp (55.3 ms/token on a 3B model) and SGLang (17.0 ms) with the CPU executing every layer (research 12 §6). A GPU spilling into this memory pays it only for the share of each token read from RAM (lib/perf/decode.ts refinement 3). System RAM.active_share 0.7924 — placement per_token_overhead_ms 35.67706871212236 ms — derived: 0.208 × 1.58 ms (GPU VRAM) + 0.792 × 44.61 ms (System RAM) = 35.68 ms bytes_per_token_gb 18.4738 GB MoE: only active parameters are read per token (SPEC §3). Effective bits include embeddings and higher-precision tensors in mixed quants. Assumes active weights are quantized like the file average. KV cache is re-read on every token (DECISIONS D-008). Pools are read sequentially per token, so per-pool times add (harmonic combination). Per-token overhead on a split placement is DERIVED, not measured: each pool's own single-pool figure is charged in proportion to the share of each token it serves, because the fixed cost is per operation and the split puts that share of the operations on each device. Calibrated with that pool serving the whole model (D-013): GPU VRAM (1.58 ms), System RAM (44.61 ms). Batch size 1 (single user, interactive).	21.1 s (estimate) ESTIMATEQwen3-VL 235B-A22B Instruct time to first token on Hybrid · Radeon AI PRO R9700 32GB (32358-01-20G) + EPYC 9015 · 8× Server Premier KSM64R52BD4-128MB 128GB DDR5-6400 RDIMMArithmetic that also depends on a tunable assumption, listed below. ttft_s = prompt_tokens / prefill_tok_s = 32,768 / 1,549.72 = 21.1 s prompt_tokens 32768 tokens prefill_tok_s 1549.72 tok/s No published prefill measurement for this hardware; scaled from peak compute (estimate). Model q_dim unknown — attention FLOPs omitted, so long-prompt TTFT is understated.	$35,377 (derived) DERIVEDHybrid · Radeon AI PRO R9700 32GB (32358-01-20G) + EPYC 9015 · 8× Server Premier KSM64R52BD4-128MB 128GB DDR5-6400 RDIMM observed priceArithmetic on observed inputs only. No tunable assumption enters it. observed_price = Σ (cheapest observed unit price × quantity) $598 + $839 + $3,928 × 8 + $1,604 + $70 + $180 + $660 = $35,377 AMD EPYC 9015 59844 ¢ — SabrePC · new · tier 2/sabrepc_ngstate ASRock Rack GENOAD8QM3-2T/BCM 83900 ¢ — Newegg · new · tier 3/newegg_realtime Kingston Server Premier KSM64R52BD4-128MB 128GB DDR5-6400 RDIMM 392811 ¢ — SabrePC · new · tier 2/sabrepc_ngstate Sapphire Radeon AI PRO R9700 32GB (32358-01-20G) 160431 ¢ — SabrePC · new · tier 2/sabrepc_ngstate ARCTIC Freezer 4U-SP5 6999 ¢ — ARCTIC · new · tier 1/arctic_jsonld Phanteks Enthoo Pro 2 Server Edition 17999 ¢ — Newegg · new · tier 3/newegg_realtime Seasonic PRIME PX-2200 65999 ¢ — Newegg · new · tier 3/newegg_realtime Listed price only: shipping and sales tax are NOT in this figure. One price per part — the cheapest monitored, non-quarantined observation.