Alibaba releases Qwen3.8-Max as open weights, with a runnable 27B to follow
Originally announced by Qwen . We link the primary source so you can read it for yourself.
Update, 14 August 2026. The weights have since been published, and the details are now settled. See our follow-up, Qwen3.8 arrives in the open: the Qwen3.8-27B is out under Apache 2.0 and is multimodal, while the flagship shipped text-only under a custom licence. The report below stands as it was written on 3 August.
Alibaba’s Qwen team has announced Qwen3.8-Max, described as its most capable model to date, and said that both it and a smaller Qwen3.8-27B model will be released as open weights. For a local AI audience, the smaller model is the one that matters, and a couple of important details are not yet settled.
The news
Qwen3.8-Max is a very large mixture-of-experts model, reported at 2.4 trillion total parameters with around 95 billion active per token, and a 1-million-token context window. It is available now through Alibaba’s own cloud and API, with open weights said to be arriving on Hugging Face and ModelScope. A second, far smaller checkpoint, Qwen3.8-27B, is also going open-weight, and that is the one sized for ordinary on-premise GPU hardware.
Two caveats matter, and we would rather state them plainly than gloss over them. As of writing, neither model’s weights are on Hugging Face yet, and no licence has been named. Qwen’s recent models have shipped under Apache 2.0, but until the licence for this release is confirmed, that is an assumption, not a fact. The performance figures in the announcement are Alibaba’s own, and we treat them as claims pending independent testing.
Why it matters
There is an honest gap between the headline and the practical story here. A 2.4-trillion-parameter model is server territory, well beyond any single machine most people own, so for the audience running models at home or on-premise, Qwen3.8-Max itself is more a statement of intent than something you will run. The Qwen3.8-27B model is the release to watch, because a 27B model fits comfortably on a 24GB card at a sensible quantisation.
The two things that will decide how much this release matters for local AI are both currently unknown: the licence, which determines whether you can use the 27B model commercially without conditions, and the actual weights, which determine when you can run it at all. Both are said to be coming.
The history
This continues one of the most active release lines in open-weight AI. It follows Qwen3.5 in February and Qwen3.6 in April, and the announcement says Qwen3.8-Max is built on the Qwen3.5 architecture. The cadence itself is part of the story: Qwen has shipped frequently and, so far, permissively, which is much of why the family has become a default for people running models on their own hardware.
What to watch
We will add Qwen3.8 entries to the model catalogue once the weights are published and the licence is confirmed, and not before, because the memory requirements and commercial terms are exactly the details our catalogue exists to get right. In the meantime, the current Qwen line, such as Qwen3 32B and the mixture-of-experts Qwen3 30B-A3B, remains a strong local choice, and can be run today through Ollama or llama.cpp.