Apache 2.0 won the permissive-licence race, and Meta is the outlier
Gemma 4, Qwen 3.5, Mistral Large 3 and Yi all ship Apache 2.0. DeepSeek V4 ships MIT. Meta still ships its own licence. Convergence on standard terms is a bigger deal than any single release.
Across the current open-weight field, Gemma 4, Qwen 3.5, Mistral Large 3 and Yi ship under Apache 2.0, DeepSeek V4 ships MIT, and Meta continues to ship Llama 4 under the Llama Community Licence.
Licence convergence is the least glamorous and most useful thing that can happen to a field. Apache 2.0 and MIT are terms that corporate counsel already have an answer for. A bespoke community licence, however reasonable its text, requires a fresh review, and legal review is where an evaluation quietly dies inside a large company.
The competitive effect is that the standard-licence models get tried and the bespoke-licence model gets scheduled. That is not a judgement about quality, and Llama's distribution is large enough to absorb the friction — but friction compounds, and the labs that removed it did so on purpose.
The remaining ambiguity is what "open" means when the weights are permissive and the training data is not disclosed. Apache 2.0 on a model whose corpus is undocumented gives you freedom to use and no ability to audit. The licence race has been won; the transparency question was never on the same track.
For anyone choosing a base model this quarter, the practical rule is unchanged: the licence determines how fast you can move, and only one major vendor is still asking you to move slowly.
Presenc AI — Open-Weight License Landscape 2026 → · jamesm.blog — The State of Open-Weight Models in 2026: Llama, Qwen, Mistral, DeepSeek →