Frontier open weights ship under a revenue-share licence
Qwen3.8-Max's weights are downloadable under a new revenue-share licence rather than Apache 2.0 or MIT. It is a third category — not permissive, not proprietary — and it arrives two weeks after Meta abandoned its bespoke licence for Apache 2.0.
The industry spent the year converging on permissive licensing. Meta gave up its community licence for Apache 2.0. Z.ai and DeepSeek were already there. A revenue-share grant on a frontier-scale model is a move in the opposite direction, and it deserves to be named as one rather than filed under "open".
Revenue-share sits in a genuinely new position. The weights are inspectable, modifiable and self-hostable — real properties that a closed API does not offer. But a licence that claims a portion of downstream revenue is a commercial relationship, and it reintroduces exactly the legal review that permissive licensing exists to eliminate.
Whether it is workable depends on details that determine everything and are easy to skim past: what counts as revenue, what threshold triggers the obligation, whether internal use is captured, and how it interacts with the ordinary case of a model buried three layers inside a product. Those questions are why the Llama community licence failed to satisfy corporate counsel, and they do not become easier when the answer is a percentage.
The honest framing is that "open weights" has stopped being one thing. Apache 2.0, community licences with user-count tripwires, and now revenue-share are three different products wearing the same word — and only one of them lets a lawyer say yes without a meeting.
explainX.ai — Qwen3.8-Max Open Weights Are Live (August 2026) → · CNBC — Meta to open source its most powerful AI model →