Alibaba announced Qwen3.8-Flash on 27 August as a preview of the architecture behind Qwen4. The weights really are free to download, but they are named Qwen3.8-Flash-Next and come under a community licence that requires a separate agreement if you build a coding or office assistant.
- 125 billion parameters in the main model, 51 billion additional N-gram embeddings, 6 billion active per token.
- The press release and the repository describe two different things: Qwen3.8-Flash-Next is the weights, Qwen3.8-Flash is the paid version in their cloud.
- The licence is Qwen Community 1.0, not Apache 2.0: model-as-a-service or a coding assistant needs separate permission.
I opened the repository first and the press release second. Turns out the order matters.
On 27 August Alibaba announced Qwen3.8-Flash - a multimodal mixture-of-experts model they say holds its own against far more expensive rivals, and an early look at the architecture behind Qwen4. The numbers are good: 125 billion parameters in the main model, 51 billion more for N-gram embeddings, and only 6 billion doing work on each token. Native context of 262 thousand tokens, stretchable to a million.
The weights, however, are not called that. What sits on Hugging Face is Qwen3.8-Flash-Next and its FP8 sibling, uploaded three days earlier. Their own model card explains it honestly: Qwen3.8-Flash is the official version in their paid cloud, built on Next, with a million tokens of context by default and built-in tools. So you download one thing, and the numbers in the announcement belong to another.
What the licence actually says
I read all of it, because for us that is a required step before any outside model goes near real work. The community licence grants the right to use, modify, distribute and sell - and then attaches two conditions.
The first is soft: cross 100 million monthly users or 20 million dollars in monthly revenue, and the model's name has to appear prominently in your product's interface. The second is not soft. If your business is giving access to models as a service, or building an assistant for coding and office work, you need separate permission from Qwen before any commercial use. Internal use stays free, as long as you do not expose the model, its outputs or its capabilities to a third party.
I am not writing this to scold them. The clause aimed at coding assistants is understandable - they sell exactly that product and would rather not fund their own competition. I am writing it because if you build software and read a headline with the words open weights, it is easy to decide you are free. A downloaded model is not the same as permitted work.
Download it, try it, measure it - nobody is asking about that. But before it goes under a paid product, open the LICENSE file in their repository and work out which of the two paragraphs you fall into.