Alibaba (Qwen) · 2026-08-12 · major
Qwen3.8-2.4T-A95B — the open-weights core of Qwen3.8-Max hits Hugging Face
Qwen3.8-2.4T-A95B is now a public download on Hugging Face, the first time Qwen has opened the weights of a Max-class model. The 2.4T mixture-of-experts fires 95B parameters per token and ships with an FP8 build.
Qwen's 2.4T Max-class mixture-of-experts is now a public download, in BF16 and FP8.
Quick facts
| Maker | Alibaba (Qwen) |
|---|---|
| Size | 2.4T total, 95B active per token |
| Experts | 512 total; 10 routed + 1 shared per token |
| Context window | 262,144 native, up to 1,010,000 |
| License | qwen3.8-max (Qwen's own licence) |
| Modality | Text only, thinking mode always on |
| Formats | BF16 and FP8 checkpoints |
Benchmarks
What is it?
Qwen3.8-2.4T-A95B is the open-weights checkpoint behind Qwen3.8-Max, published on Hugging Face on 12 August 2026 along with an FP8 build. It is the first time Qwen has released the weights of a Max-class model instead of keeping the flagship tier behind its API. The checkpoint is text only and always reasons.
How does it work?
A 92-layer stack alternates Gated DeltaNet linear attention with gated full attention, and each block ends in a mixture-of-experts layer. Of 512 experts, 10 routed plus 1 shared fire per token, which is how 2.4T total parameters cost only 95B of compute per step. The FP8 build uses fine-grained quantization with a block size of 128, and Qwen says its results are nearly identical to the original.
Why does it matter?
Until today the 2.4T flagship was API-only, so outsiders could not audit its published scores, fine-tune it, or keep their data off Alibaba's servers. Downloadable weights change all three: teams can serve the model themselves on vLLM, SGLang or TokenSpeed, and community quantizations for llama.cpp, LM Studio, Jan and Ollama are already listed on the card. The catch is the licence, which is Qwen's own rather than Apache-2.0.
Who is it for?
teams self-hosting frontier open-weight models
Frequently asked questions
- Is Qwen3.8-2.4T-A95B free for commercial use?
- Qwen3.8-2.4T-A95B does not ship under Apache-2.0 like most of Qwen's open family. The model card lists its licence as qwen3.8-max, a licence Qwen wrote for this release. Anyone planning to build a product on the weights should read that licence file first rather than assume the usual Apache terms carry over.
- How is Qwen3.8-2.4T-A95B different from Qwen3.8-Max?
- Qwen3.8-Max is the official service version built on Qwen3.8-2.4T-A95B, and it adds features the open checkpoint does not have. The model card says Max supports vision input and non-thinking mode. The downloadable Qwen3.8-2.4T-A95B is text only, and reasoning cannot be switched off.
- Can you turn off thinking mode in Qwen3.8-2.4T-A95B?
- No. Qwen3.8-2.4T-A95B requires thinking mode for every interaction, and each reply begins with reasoning wrapped in a think block. What you can change is how hard it thinks: the card exposes reasoning effort levels of xhigh, which is the default, plus medium and low.
- How much of the 1M context can Qwen3.8-2.4T-A95B spend on output?
- Qwen3.8-2.4T-A95B handles 262,144 tokens natively and stretches to 1,010,000. Inside that window the card caps reasoning at 262,144 tokens and the final answer at 131,072 tokens, so a long-context job needs budget for both the visible reply and the hidden thinking that precedes it.
Try it
vllm serve "Qwen/Qwen3.8-2.4T-A95B"