Qwen · Chat and reasoning · 2.4T parameters · 95B active · Qwen3.8 licence (revenue threshold $50M)
Qwen3.8 2.4T A95B
Needs 2,880 GB at FP8 / INT8: more than eight GPUs of any board type. Multi-node only.
Parameters2.4T95B active per token
Context256k tokensmaximum published
Checkpoint4,892 GBBF16, as published
LicenseQwen3.8 licence (revenue threshold $50M)check before commercial use
ReleasedAug 2026documented date
What is documented about Qwen3.8 2.4T A95B
news checked 2026-10-27Open weights since August 2026 under a specific licence with a $50M annual-revenue threshold. Sized at about 1.2 TB in 4-bit: 8× B300 or 8× MI355X as a minimum, two HGX B200 nodes or more for interactive use, NVL72 or multi-node for production.
- https://huggingface.co/Qwen/Qwen3.8-2.4T-A95B
- https://huggingface.co/Qwen/Qwen3.8-2.4T-A95B/raw/main/config.json
- https://huggingface.co/api/models/Qwen/Qwen3.8-2.4T-A95B?blobs=true
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 | 4,892 GB | 5,871 GB | 5,873 GB | 5,884 GB | 5,921 GB | 5,970 GB |
| FP8 / INT8estimated | 2,400 GB | 2,880 GB | 2,882 GB | 2,893 GB | 2,930 GB | 2,979 GB |
| INT4 / MXFP4estimated | 1,440 GB | 1,728 GB | 1,730 GB | 1,741 GB | 1,778 GB | 1,827 GB |
| KV cacheone sequence, FP16 cache, included above | — | — | 2 GB | 13 GB | 50 GB | 99 GB |
Cross-checked on Hugging Face Oct 28, 2026: 2446.2B parameters in the safetensors headers, 4892.4 GB of weight files, attention layout from config.json (92 layers, 4 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 = 4096 bytes per token per layer × 92 full-attention layers (FP16 cache, one sequence).config.json ↗
Smallest configuration per GPU · 109 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 |
|---|---|---|---|---|---|
| No configuration up to 8 GPUs of one type fits 2880 GB at FP8 / INT8: multi-node only. | |||||
Qwen3.8 2.4T A95B · related