Model releases · 12 August 2026
The largest model Qwen has ever published, released first at the top of the range rather than the bottom, and under a bespoke licence instead of the Apache 2.0 that covered every size of Qwen3.5.
| Released by | Alibaba (Qwen team) |
|---|---|
| Published | 12 August 2026 |
| Parameters | 2.4T total, 95B activated per token |
| Architecture | Mixture-of-Experts (Qwen3_5MoeForCausalLM, qwen3_5_moe_text) |
| Layers / hidden size | 92 / 8192 |
| Experts | 512 routed, top-10 per token |
| Attention | 64 Q heads / 4 KV heads, plus linear attention heads (128 V, 16 QK) |
| Context length | 262,144 tokens (1M by default in the hosted Max version) |
| Vocabulary | 248,320 tokens |
| Licence | qwen3.8-max — custom, not Apache 2.0 |
| Weight format | safetensors — 4,892 GB across 213 shards; FP8 build 2,496 GB |
| Hosted at | Qwen/Qwen3.8-2.4T-A95B on Hugging Face |
Qwen3.8-2.4T-A95B is the open-weight base of Qwen3.8-Max, Alibaba's hosted flagship. The repository ships the post-trained model in Transformers format, and Qwen states the artifacts work with vLLM, SGLang and similar serving stacks.
The hosted Max version carries features this release does not: vision input, a non-thinking mode, a 1M-token context by default, and built-in tools. What you can download is the text model.
Every size of Qwen3.5 — from 0.8B up to the 397B sparse model — shipped under Apache 2.0. That consistency is a large part of why the family became the default base for other people's work, from French language specialists to Microsoft's own browser agent.
Qwen3.8-2.4T-A95B ships under a bespoke qwen3.8-max licence instead. Anyone planning to build on it should read the terms directly rather than assume the family's previous permissiveness carries over.
Whether the smaller Qwen3.8 sizes, when they appear, return to Apache 2.0 or follow the flagship is the open question that matters most for downstream work.
The model type string is qwen3_5_moe_text — Qwen3.8 is built on the Qwen3.5 architectural foundation rather than introducing a new one. Ninety-two layers at hidden size 8192, with 512 routed experts and the top 10 selected per token.
The attention configuration is unusual: 64 query heads against 4 key-value heads, alongside separate linear attention heads (128 for V, 16 for QK). Qwen also exposes reasoning_effort to tune reasoning depth and preserve_thinking to retain reasoning context across turns.
Only the 2.4T model was published. The 27B size named in Qwen's own pre-release messaging is not in the repository, and no smaller variants exist at the time of writing. For anyone whose interest in Qwen is the 0.8B to 4B band, this release changes nothing yet.
4,892 GB of weights across 213 shards, or 2,496 GB for the FP8 build. This is multi-node infrastructure, not a workstation model.
| Runtime | Supported | Notes |
|---|---|---|
| transformers | Yes | Reference format published by Qwen. |
| vLLM / SGLang | Yes | Explicitly named as compatible in the model card. |
| llama.cpp | No | No GGUF published, and no community conversion at a size that would be practical. |
| MLX | No | Not published. |
| Official quantised build | Yes | An FP8 repository is published alongside the bfloat16 weights. |
qwen3.8-max licence have not been widely analysed.2.4 trillion in total with 95 billion activated per token, across 512 routed experts with the top 10 selected per token.
No. Qwen3.8-2.4T-A95B ships under a custom qwen3.8-max licence. Every size of Qwen3.5 was Apache 2.0, so this is a change in the family's terms.
Not at the time of writing. Only the 2.4T model has been published, despite a 27B size being named ahead of the release.
Max is the hosted official version built on these weights, adding vision input, a non-thinking mode, a 1M-token default context and built-in tools. The open release is the text model.
4,892 GB across 213 safetensors shards for bfloat16. The FP8 build is 2,496 GB.
OnDevice LLM is a private AI assistant that runs entirely on your iPhone — no account, no cloud, and nothing you type leaves the device.
Published 12 August 2026