Four open-weight labs, four incompatible licences
Kimi ships under the Kimi K3 License. DeepSeek publishes its own model terms. Qwen varies the licence per model. Only some releases are Apache 2.0. "Open weights" has stopped describing a single thing.
A survey of the current open-weight field turns up four different legal regimes across four labs. Kimi K3 carries a custom Kimi K3 License with conditions attached to some large commercial services. DeepSeek publishes weights under its own model terms. Individual Qwen models carry different licences from one another. And a subset of releases across the field are genuinely Apache 2.0.
This is not a complaint about permissiveness. It is an observation that the phrase has stopped carrying information. Two models both described as open-weight can differ on commercial use, on scale thresholds, on redistribution, and on whether you may train on the outputs — and a reader comparing benchmark scores sees none of that.
The practical consequence lands on anyone building a product rather than a demo. Swapping the model underneath is a legal review, not a config change, and the review is different for every candidate. That is a real switching cost, and it is one the benchmark tables do not price.
There is a version of this that resolves into a small number of standard forms, the way software licensing eventually did. There is another version where every lab keeps its own terms permanently, because bespoke terms are exactly how you capture value from something you gave away. Nothing in the current field distinguishes the two.
GEO Toolbox — Chinese AI Models Compared: DeepSeek, Qwen, GLM, Kimi (2026) → · GitHub — open-weight-models — curated list with commercially exploitable licenses →