Qwen · Chat and reasoning · 403B parameters · 17B active · apache-2.0
Qwen3.5 397B A17B
Needs 484 GB at FP8 / INT8. The smallest real configuration today is 2× MI325X at Cyfuture AI, $6.21 per hour, about $4,536 a month at 730 hours.
Parameters403B17B active per token
Context256k tokensmaximum published
Checkpoint807 GBBF16, as published
Licenseapache-2.0check before commercial use
ReleasedFeb 16, 2026documented date
Smallest configuration · FP8 / INT8
2× MI325X512 GB VRAM · 94% used
- Per hour
- $6.21
- Per month
- $4,536
- Cheapest at
Cyfuture AI
What is documented about Qwen3.5 397B A17B
news checked 2026-10-27Released 2026-02-16 under Apache 2.0. About 220 GB in NVFP4; a vLLM forum post reports 50.5 tok/s single-user on 4× RTX PRO 6000 (Marlin W4A16, TP=4).
Memory · 3 precisions · 5 context lengthsHow much VRAM it needs
Weights plus runtime overhead, then the KV cache that grows with the context you serve.
| Precision | Weights | Weights + headroom | 4k context | 32k context | 128k context | 256k context |
|---|---|---|---|---|---|---|
| BF16published checkpoint | 807 GB | 969 GB | 969 GB | 973 GB | 985 GB | 1,001 GB |
| FP8 / INT8estimated | 403 GB | 484 GB | 485 GB | 488 GB | 500 GB | 516 GB |
| INT4 / MXFP4estimated | 242 GB | 291 GB | 291 GB | 295 GB | 307 GB | 323 GB |
| KV cacheone sequence, FP16 cache, included above | — | — | 1 GB | 5 GB | 17 GB | 33 GB |
Cross-checked on Hugging Face Oct 28, 2026: 403.4B parameters in the safetensors headers, 806.8 GB of weight files, attention layout from config.json (60 layers, 2 KV heads × 256). Method: weights as published at the native precision, verified on Hugging Face 2026-10-28; other precisions estimated at 1 (8-bit) and 0.6 (4-bit) bytes per parameter; × 1.2 runtime headroom; KV cache = 2048 bytes per token per layer × 60 full-attention layers (FP16 cache, one sequence).config.json ↗
Smallest configuration per GPU · 110 GPUs · cheapest eligible on-demand offer per GPU count, every provider on fileWhat it costs to run, GPU by GPU
The fewest GPUs of one type whose total memory fits the model at short context, priced from the cheapest eligible on-demand offer for exactly that count (45-day window, newest reading per provider), with the next providers beside it. Monthly = 730 hours.
| GPU | Smallest config | VRAM used | $ / hour | $ / month | Cheapest at |
|---|---|---|---|---|---|
| MI325XBest fit | 2× 256 GB512 GB total | 94% tight | $6.21 | $4,536 | Cyfuture AInot scored · read Nov 2 · receipt |
| RTX PRO 6000 | 8× 96 GB768 GB total | 63% | $7.78 | $5,676 |  UpCloudTrust 75 · read Nov 2 · receipt · also Vast.ai $9.58, Lium $12.80 |
| Gaudi 2 | 8× 96 GB768 GB total | 63% | $8.27 | $6,036 | Cyfuture AInot scored · read Nov 2 · receipt |
| A100 80GB | 8× 80 GB640 GB total | 76% | $8.54 | $6,231 |  Vast.aiTrust 78 · read Nov 2 · receipt · also Edgevana $9.82, GPU.ai $10.00 |
| H200 NVL | 4× 141 GB564 GB total | 86% | $10.67 | $7,788 | Vast.aiTrust 78 · read Nov 2 · receipt · also IONOS $13.04, Geodd $13.56 |
| MI300X | 4× 192 GB768 GB total | 63% | $11.31 | $8,257 | Cyfuture AInot scored · read Nov 2 · receipt · also Hot Aisle $11.96, AtmosCompute $13.38 |
| B300 | 2× 288 GB576 GB total | 84% | $11.65 | $8,506 | UpCloudTrust 75 · read Nov 2 · receipt · also FPT AI Factory $13.98, Vast.ai $17.00 |
| H100 | 8× 80 GB640 GB total | 76% | $13.67 | $9,982 |  GcoreTrust 80 · read Nov 2 · receipt · also Airon AI $15.04, UpCloud $16.00 |
| H200 | 4× 141 GB564 GB total | 86% | $14.76 | $10,775 |  Theta EdgeCloudnot scored · read Nov 2 · receipt · also Canopy Wave $16.00, Hugging Face $20.00 |
| B200 | 4× 180 GB720 GB total | 67% | $16.46 | $12,013 | UpCloudTrust 75 · read Nov 2 · receipt · also Deep Infra $19.56, Lambda $27.16 |
| H100 NVL | 8× 94 GB752 GB total | 64% | $16.48 | $12,030 |  GPU.ainot scored · read Nov 2 · receipt · also Spheron $21.97, AtmosCompute $35.73 |
| MI355X | 8× 288 GB2304 GB total | 21% | $23.60 | $17,228 |  TensorWaveTrust 64 · read Nov 2 · receipt |
| GB300 | 2× 288 GB576 GB total | 84% | $26.11 | $19,060 |  VerdaTrust 81 · read Nov 2 · receipt |
| B200 CC | 4× 180 GB720 GB total | 67% | $36.19 | $26,419 | VerdaTrust 81 · read Nov 2 · receipt |
| B300 CC | 4× 288 GB1152 GB total | 42% | $46.31 | $33,806 | VerdaTrust 81 · read Nov 2 · receipt |
| A800 80GB PCIe | 8× 80 GB640 GB total | 76% | $50.80 | $37,088 |  Volcano Enginenot scored · read Nov 2 · receipt |
Throughput estimates · Oct 4, 2026Tokens per second, and what a million costs
Chat and reasoning. Speed estimated from memory bandwidth and model size, never measured. The cost per 1M output tokens is an estimate at the prices read on Oct 4, 2026; it is not recomputed here.
| GPU | Cards | Precision | tok/s per user | per 1M output tokens · estimate |
|---|---|---|---|---|
| MI300X | 4 | FP8 | 42.8 | $1.77 |
| H200 | 4 | FP8 | 55.3 | $2.19 |
| B200 | 4 | FP8 | 74.2 | $2.35 |
| H100 SXM | 8 | FP8 | 45.5 | $3.07 |
| RTX PRO 6000 | 8 | FP8 | 28.9 | $4.43 |
| H100 PCIe | 8 | FP8 | 27.1 | $4.88 |
cards, precision and speed estimated from memory bandwidth and model size; cost recomputed by the site from verified prices · cheapest per output at the prices of Oct 4, 2026: MI300XWhy estimates, and their limits →
Qwen3.5 397B A17B · related