The MIT-licence frontier
The most permissive licences in software are now attached to models close to the frontier. That is a strategic choice, not an oversight.
DeepSeek V4 Flash shipped MIT-licensed, published to the hour, superseding its preview checkpoint. Qwen 3.6-27B stays downloadable while Alibaba's current generation moves API-only.
MIT is more permissive than most people register
No patent grant obligations. No attribution beyond the notice. At frontier-adjacent capability that is close to public domain in practice, and it is a materially different proposition from Apache 2.0, which is itself a long way past the community licences most large releases carry.
The release cadence is the tell
A preview checkpoint replaced by a dated production build, published at a specific time on a specific day. That is a software release process applied to model weights, and it implies infrastructure for shipping models the way other companies ship binaries.
Labs that treat weights as releases rather than events will out-iterate labs that treat each one as a launch.
The two-tier settlement
Downloadable weights for the previous generation, API for the current. It keeps an ecosystem alive without giving away the frontier, and several labs appear to be converging on it independently.
Whether it holds depends entirely on how fast the open tier ages. One generation behind is a reasonable settlement. Three generations behind is a marketing artefact with a licence attached.
Where this leaves the argument
OpenAI is now in the same territory, which removes the last version of the argument that open weights are someone else's strategy.
Twenty-seven billion parameters is the size that fits a workstation, which is the size that decides whether a company with data-residency constraints can use a model family at all. That is not a benchmark question. It is a question about who is allowed to participate.
Hugging Face — DeepSeek V4 Flash is now official: what changed in the 0731 build → · Hugging Face — Qwen/Qwen3.6-27B → · OpenAI — Introducing gpt-oss →