7th of 36 machines that run your plan, lowest price first.

2× Radeon AI PRO R9700 Creator 32GB (R9700 CT) · EPYC 9015 · 8× M321R4GA3BB6-CQK 32GB DDR5-4800 RDIMM

Not confirmed to run your plan: the sizes of Qwen3.8-Flash-Next's and Qwen3-ASR 1.7B's projector files (mmproj) are not recorded, so they are not counted. 2× Radeon AI PRO R9700 Creator 32GB (R9700 CT) · EPYC 9015 · 8× M321R4GA3BB6-CQK 32GB DDR5-4800 RDIMM has 0.27 GB left for them.

At least $7,462 from Newegg, SabrePC and 1 more

How it runs your setup

Tokens a second while writing an answer. Your plan asks for 4.7 or more tokens a second.

CodeGLM-5.3-Flash
First reply after reading 32,768 tokens: 16 seconds
15 tokens a secondKeeps up with you
ImagesQwen3.8-Flash-Next
First reply after reading 32,768 tokens: 57 seconds
5.5 tokens a secondFine to wait
SpeechQwen3-ASR 1.7B
First reply after reading 30,000 tokens: 1.5 seconds
90 tokens a secondInstant

Memory with all 3 loaded306 GB of 305 GB

This machine leaves 305 GB of its 320 GB for AI. 4.7 GB of what your models need keeps your conversations cached. That counts measured memory only: the sizes of Qwen3.8-Flash-Next's and Qwen3-ASR 1.7B's projector files (mmproj) are not recorded.

Follow-up repliesOnly new text is readEvery model stays loaded with its conversation cached, so a follow-up reads only what you add.

Show the work

Each token means reading the model's weights from memory, so speed follows memory bandwidth; the first reply follows how fast the machine reads the prompt. Estimates are calibrated against published benchmarks; your app and its settings will move them a little.

GLM-5.3-Flash

Writing speed: estimate — Arithmetic 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: 3.2 GB ÷ (640 GB/s × 0.48) + System RAM: 5.549 GB ÷ (307.2 GB/s × 0.7) + 28.87 ms fixed per-token overhead (share-weighted) → 15.36 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: 0.3657 — 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: 0.6343 — placement
per_token_overhead_ms: 28.872107066297204 ms — derived: 0.366 × 1.58 ms (GPU VRAM) + 0.634 × 44.61 ms (System RAM) = 28.87 ms
bytes_per_token_gb: 8.7487 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).

First reply: estimate — Arithmetic that also depends on a tunable assumption, listed below.

ttft_s = prompt_tokens / prefill_tok_s

= 32,768 / 1,996.44 = 16.4 s

prompt_tokens: 32768 tokens
prefill_tok_s: 1996.44 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.

Qwen3.8-Flash-Next

Writing speed: estimate — Arithmetic 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: 19.069 GB ÷ (640 GB/s × 0.48) + System RAM: 20.706 GB ÷ (307.2 GB/s × 0.7) + 23.98 ms fixed per-token overhead (share-weighted) → 5.48 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: 0.4794 — 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: 0.5206 — placement
per_token_overhead_ms: 23.980284367226076 ms — derived: 0.479 × 1.58 ms (GPU VRAM) + 0.521 × 44.61 ms (System RAM) = 23.98 ms
bytes_per_token_gb: 39.7749 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).

First reply: estimate — Arithmetic that also depends on a tunable assumption, listed below.

ttft_s = prompt_tokens / prefill_tok_s

= 32,768 / 577.57 = 56.7 s

prompt_tokens: 32768 tokens
prefill_tok_s: 577.57 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.

Qwen3-ASR 1.7B

Writing speed: estimate — Arithmetic that also depends on a tunable assumption, listed below.

tok_s = 1 / (bytes_per_token / (bandwidth × efficiency) + overhead)

GPU VRAM: 5.274 GB ÷ (640 GB/s × 0.86) + 1.58 ms fixed per-token overhead → 89.59 tok/s

GPU VRAM.bandwidth_gbs: 640 GB/s — component spec
GPU VRAM.efficiency: 0.86 — 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: 5.2741 GB

Dense: every parameter is read per token.
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).

First reply: estimate — Arithmetic that also depends on a tunable assumption, listed below.

ttft_s = prompt_tokens / prefill_tok_s

= 30,000 / 19,426.65 = 1.5 s

prompt_tokens: 30000 tokens
prefill_tok_s: 19426.65 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.

Memory

Every model loaded: estimate — Arithmetic that also depends on a tunable assumption, listed below.

Σ required_gb over every model in the stack

GLM-5.3-Flash 169.74 + Qwen3.8-Flash-Next 129.49 + Qwen3-ASR 1.7B 5.78 = 305 GB

GLM-5.3-Flash: 169.736978848 GB
Qwen3.8-Flash-Next: 129.48808638719998 GB
Qwen3-ASR 1.7B: 5.77887773952 GB

All models loaded at once (Keep all resident).
Not counted, because it is not measured yet: Qwen3.8-Flash-Next reads images through a separate projector file (mmproj); its size is not recorded yet, so the memory shown leaves it out. Qwen3-ASR 1.7B hears audio through a separate projector file (mmproj); its size is not recorded yet, so the memory shown leaves it out.

Memory for AI: derived — Arithmetic 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).

To spare: estimate — Arithmetic that also depends on a tunable assumption, listed below.

spare_gb = usable_gb − all-resident need

= 305.28 − 305 = 0.28 GB

usable_gb: 305.28 GB
need_gb: 305.00394297471996 GB

All models loaded at once (Keep all resident).
Not counted, because it is not measured yet: Qwen3.8-Flash-Next reads images through a separate projector file (mmproj); its size is not recorded yet, so the memory shown leaves it out. Qwen3-ASR 1.7B hears audio through a separate projector file (mmproj); its size is not recorded yet, so the memory shown leaves it out.
So the real room to spare is smaller than shown, by the size of what is not counted: the sizes of Qwen3.8-Flash-Next's and Qwen3-ASR 1.7B's projector files (mmproj) are not recorded.

GLM-5.3-Flash: 170 GB — required_gb = 156.82 + 0.37 + 12.55 = 169.74 GB
Qwen3.8-Flash-Next: 130 GB — required_gb = 119.15 + 0.81 + 9.53 = 129.49 GB
Qwen3-ASR 1.7B: 5.8 GB — required_gb = 2.17 + 3.44 + 0.17 = 5.78 GB

What's inside

Processor: AMD EPYC 9015 — 8 cores, 12 memory channels
Graphics card: 2× ASRock Radeon AI PRO R9700 Creator 32GB (R9700 CT) — 32 GB GDDR6 (ECC, Linux only), 640 GB/s
Motherboard: ASRock Rack GENOAD8QM3-2T/BCM
Memory: 8× Samsung M321R4GA3BB6-CQK 32GB DDR5-4800 RDIMM — 32 GB DDR5 RDIMM at 4,800 MT/s
Power supply: Seasonic PRIME PX-2200 — 2,200 W, 80 PLUS Platinum
Cooler: ARCTIC Freezer 4U-SP5
Case: Phanteks Enthoo Pro 2 Server Edition
Kind: Build your own — parts to buy and put together

Set it up

Three steps once it arrives. The apps named here are free; each model's license is listed with it.

1
Step 1: Install an app that runs models
- llama.cpp (opens in a new tab) · LM Studio (opens in a new tab) · Ollama (opens in a new tab)For GLM-5.3-Flash, Qwen3.8-Flash-Next, Qwen3-ASR 1.7B: GGUF files run in llama.cpp, and apps built on it, such as LM Studio and Ollama, open them too. Check each model's page for the app versions that support it.
2
Step 2: Download your 3 models
- GLM-5.3-Flash · Code157 GB
  unsloth/GLM-5.3-Flash-GGUF — UD-IQ4_XS, 5 files (opens in a new tab)mit, a permissive license.UD-IQ4_XS is split into 5 files; download all of them from the repository.
- Qwen3.8-Flash-Next · Images120 GB
  lmstudio-community/Qwen3.8-Flash-Next-GGUF — Q4_K_M, 3 files (opens in a new tab)other — read its terms before using it for work.To read images, a GGUF model also needs its mmproj projector file from the same repository. Its size is not recorded here, so it is not in the download total or the memory figures.Q4_K_M is split into 3 files; download all of them from the repository.
- Qwen3-ASR 1.7B · Speech2.2 GB
  Qwen3-ASR-1.7B-Q8_0.gguf (opens in a new tab)apache-2.0, a permissive license.To hear audio, a GGUF model also needs its mmproj projector file from the same repository. Its size is not recorded here, so it is not in the download total or the memory figures.
279 GB to download
Download size: observed — Measured or published, passed through unchanged — not computed here.
download_gb = Σ file_size_gb of the files listed
GLM-5.3-Flash 156.82 + Qwen3.8-Flash-Next 119.15 + Qwen3-ASR 1.7B 2.17 = 278.14 GB
GLM-5.3-Flash
156.8221112 GB — Hugging Face file listing
Qwen3.8-Flash-Next
119.15072224 GB — Hugging Face file listing
Qwen3-ASR 1.7B
2.165034944 GB — Hugging Face file listing
Step 3: Keep all 3 loaded
Keep all 3 loaded (GLM-5.3-Flash for Code; Qwen3.8-Flash-Next for Images; Qwen3-ASR 1.7B for Speech) so each conversation stays cached and follow-ups only read what is new.

Price

At least$7,462

Newegg, SabrePC and 1 more Out of stock

Stale · checked 29 d ago

via access tier 2, route seed:research; access tier 2, route sabrepc_ngstate; access tier 3, route newegg_realtime; access tier 1, route arctic_jsonld (landed, before sales tax) · Newegg, SabrePC, ARCTIC don't show shipping until checkout.

Price before sales tax — add a ZIP code to include it.

To run, a year

Unavailable — This machine's draw at the wall is not known: 4 power figures it needs are not published. No idle draw is published for any of its 6 parts. The duty cycle spends 18 h a day idle, and an unknown idle draw is not counted as zero.

Show the work

Landed price: derived — Arithmetic on observed inputs only. No tunable assumption enters it.

rig_landed = Σ part landed_total

$598 + $3,400 + $839 + $1,714 + $660 + $70 + $180 = $7,462

AMD EPYC 9015: 59844 ¢ — SabrePC · ship
ASRock Radeon AI PRO R9700 Creator 32GB (R9700 CT): 339998 ¢ — Newegg · ship
ASRock Rack GENOAD8QM3-2T/BCM: 83900 ¢ — Newegg · ship
Samsung M321R4GA3BB6-CQK 32GB DDR5-4800 RDIMM: 171416 ¢ — SabrePC · ship
Seasonic PRIME PX-2200: 65999 ¢ — Newegg · ship
ARCTIC Freezer 4U-SP5: 6999 ¢ — ARCTIC · ship
Phanteks Enthoo Pro 2 Server Edition: 17999 ¢ — Newegg · ship

Some part prices are older than 36 h.
LOWER BOUND: 7 of 7 parts come from a retailer that does not state a delivery charge (SabrePC, Newegg, Newegg, SabrePC, Newegg, ARCTIC, Newegg), so the rig costs at least this and never less.

Why there is no running cost

This machine's draw at the wall is not known: 4 power figures it needs are not published. No idle draw is published for any of its 6 parts. The duty cycle spends 18 h a day idle, and an unknown idle draw is not counted as zero.
Electrical feasibility cannot be computed exactly: no load or TDP watts for ASRock Rack GENOAD8QM3-2T/BCM, Samsung M321R4GA3BB6-CQK 32GB DDR5-4800 RDIMM, ARCTIC Freezer 4U-SP5, Phanteks Enthoo Pro 2 Server Edition.
No idle draw is published for AMD EPYC 9015, ASRock Radeon AI PRO R9700 Creator 32GB (R9700 CT), ASRock Rack GENOAD8QM3-2T/BCM, Samsung M321R4GA3BB6-CQK 32GB DDR5-4800 RDIMM, ARCTIC Freezer 4U-SP5, Phanteks Enthoo Pro 2 Server Edition.

Missing: ASRock Rack GENOAD8QM3-2T/BCM: load_watts or tdp_watts, Samsung M321R4GA3BB6-CQK 32GB DDR5-4800 RDIMM: load_watts or tdp_watts, ARCTIC Freezer 4U-SP5: load_watts or tdp_watts, Phanteks Enthoo Pro 2 Server Edition: load_watts or tdp_watts, AMD EPYC 9015: idle_watts, ASRock Radeon AI PRO R9700 Creator 32GB (R9700 CT): idle_watts, ASRock Rack GENOAD8QM3-2T/BCM: idle_watts, Samsung M321R4GA3BB6-CQK 32GB DDR5-4800 RDIMM: idle_watts, ARCTIC Freezer 4U-SP5: idle_watts, Phanteks Enthoo Pro 2 Server Edition: idle_watts.

Where to buy

Graphics card2× ASRock Radeon AI PRO R9700 Creator 32GB (R9700 CT)
Newegg (opens in a new tab)At least$3,399.98
2 × $1,699.99Delivery — shipping shown only at checkoutIn stockStale · checked 29 d agovia access tier 2, route seed:research · shipping is added at checkout
Offers from 1 of the 41 stores we check.
Memory8× Samsung M321R4GA3BB6-CQK 32GB DDR5-4800 RDIMM
SabrePC (opens in a new tab)At least$1,714.16
8 × $214.27Delivery — shipping shown only at checkoutOut of stockChecked 15 h agovia access tier 2, route sabrepc_ngstate · shipping is added at checkout
Offers from 1 of the 41 stores we check.
MotherboardASRock Rack GENOAD8QM3-2T/BCM
Newegg (opens in a new tab)At least$839.00
Delivery — shipping shown only at checkoutIn stockChecked 2 h agovia access tier 3, route newegg_realtime · shipping is added at checkout
Offers from 1 of the 41 stores we check.
Power supplySeasonic PRIME PX-2200
Newegg (opens in a new tab)At least$659.99
Delivery — shipping shown only at checkoutIn stockChecked 2 h agovia access tier 3, route newegg_realtime · shipping is added at checkout
Offers from 1 of the 41 stores we check.
ProcessorAMD EPYC 9015
SabrePC (opens in a new tab)At least$598.44
Delivery — shipping shown only at checkoutOut of stockChecked 15 h agovia access tier 2, route sabrepc_ngstate · shipping is added at checkout
Offers from 2 of the 41 stores we check.
1 other store
Central Computers (opens in a new tab)At least$699.99
Delivery — shipping shown only at checkoutOut of stockChecked 16 h agovia access tier 1, route jsonld_product · shipping is added at checkout
CasePhanteks Enthoo Pro 2 Server Edition
Newegg (opens in a new tab)At least$179.99
Delivery — shipping shown only at checkoutIn stockChecked 2 h agovia access tier 3, route newegg_realtime · shipping is added at checkout
Offers from 2 of the 41 stores we check.
1 other store
SabrePC (opens in a new tab)At least$211.86
Delivery — shipping shown only at checkoutOut of stockChecked 15 h agovia access tier 2, route sabrepc_ngstate · shipping is added at checkout
CoolerARCTIC Freezer 4U-SP5
ARCTIC (opens in a new tab)At least$69.99
Delivery — shipping shown only at checkoutIn stockChecked 16 h agovia access tier 1, route arctic_jsonld · shipping is added at checkout
Offers from 1 of the 41 stores we check.

Total for 7 partsAt least$7,461.55

The same as the price above, to the cent.

Show the work

Machine total: derived — Arithmetic on observed inputs only. No tunable assumption enters it.

rig_landed = Σ part landed_total

$598 + $3,400 + $839 + $1,714 + $660 + $70 + $180 = $7,462

AMD EPYC 9015: 59844 ¢ — SabrePC · ship
ASRock Radeon AI PRO R9700 Creator 32GB (R9700 CT): 339998 ¢ — Newegg · ship
ASRock Rack GENOAD8QM3-2T/BCM: 83900 ¢ — Newegg · ship
Samsung M321R4GA3BB6-CQK 32GB DDR5-4800 RDIMM: 171416 ¢ — SabrePC · ship
Seasonic PRIME PX-2200: 65999 ¢ — Newegg · ship
ARCTIC Freezer 4U-SP5: 6999 ¢ — ARCTIC · ship
Phanteks Enthoo Pro 2 Server Edition: 17999 ¢ — Newegg · ship

Some part prices are older than 36 h.
LOWER BOUND: 7 of 7 parts come from a retailer that does not state a delivery charge (SabrePC, Newegg, Newegg, SabrePC, Newegg, ARCTIC, Newegg), so the rig costs at least this and never less.

ASRock Radeon AI PRO R9700 Creator 32GB (R9700 CT), share of the total (rank 1): derived — Arithmetic on observed inputs only. No tunable assumption enters it.