Qwen3.8-Max's open weights landed — the license restricts exactly the product it was built for

Alibaba published open weights for Qwen3.8-Max on Hugging Face on August 12 — the mixture-of-experts model that has been promised "this week" since August 3 finally landed, under the repository Qwen/Qwen3.8-2.4T-A95B, plus an FP8 variant. It is the first time Alibaba has open-weighted a Max-class Qwen model: 2.4 trillion total parameters, 95 billion active per token across 512 experts (11 active at a time). The full-precision repository alone is 213 safetensors files totaling close to 4.9 TB — bf16 math on 2.4 trillion parameters lands almost exactly there, which is a useful sanity check if you are budgeting storage before you pull anything.

Two things are worth stopping on before you touch it.

The license is not what the last two Qwen releases trained you to expect. Qwen3.5 and Qwen3.6 shipped under plain Apache 2.0, and most coverage assumed Qwen3.8 would follow the same pattern. It does not. The repository ships under a custom "Qwen3.8-Max License," and two clauses matter for builders. First, a display requirement: if you use the model in a commercial product with more than 100 million monthly active users or more than $20 million in monthly revenue, you must prominently display the model name. Second, and more consequential: if your aggregate revenue from running the model as a hosted inference service, or from "an independent AI-powered product primarily designed for AI-assisted coding or office productivity," crosses $50 million in any trailing 12-month period, you need a separate commercial license from Qwen before you can keep using it. That second category is not a hypothetical edge case here — Qwen marketed this exact release as "a new bar for coding and cowork." The license carves out a licensing conversation for precisely the product category the model was announced for. If you are building or already running a coding assistant or office-productivity tool at that revenue scale, read the LICENSE file in the repository yourself before you plan around this model; do not carry over the Apache 2.0 assumption from Qwen3.5 or Qwen3.6.

The download is not the model the benchmarks describe. Qwen's own README lists a native 262,144-token context window, extensible to 1,010,000 via configuration, thinking mode only, and no vision input. The cloud Qwen3.8-Max API that Qwen has been marketing since August 3 serves the full multimodal model — vision input, a 1-million-token context window by default, and a non-thinking mode. Users noticed within hours: a discussion thread on the model's Hugging Face page opened the same day, pointing out that the repository is tagged with the "qwen3.8-max" license yet ships a text-only build with a quarter of the advertised context available by default. As of publication there is no official response in that thread. If the benchmarks you saw for Qwen3.8-Max were vision or long-context results, they were run against the cloud model, not the weights you can download today.

What that means in practice: if you plan to self-host this for a coding-assistant or office-productivity product, model your trailing-12-month revenue against the $50 million threshold before you build anything on it, and budget for the fact that you are getting a text-only model at a smaller default context than the one Qwen has been benchmarking publicly. At close to 4.9 TB for the full-precision weights, this also is not a single-GPU experiment the way Muse Glimmer was — Qwen's own deployment guidance points to vLLM or SGLang across a multi-GPU cluster, and community quantized builds have not had time to appear yet. The same rule that applied to Kimi K3 applies here, just after the license actually shipped instead of before it: read the license file yourself, not the coverage, and check whether the specific weights on disk match the specific product you were benchmarking against.