Kimi K3 is not MIT licensed, and a lot of people assumed it was
Moonshot shipped the current leading open-weight model under a custom licence. Coverage repeatedly described it as MIT. The gap between what a model family is marketed as and what the specific checkpoint's licence says is now a live legal exposure.
Kimi K3, released with open weights on 27 July, is the strongest open-weight model on most current comparisons — 2.8 trillion total parameters, 104 billion active. It is also shipped under a custom Moonshot licence, not MIT, despite being widely described as MIT in secondary coverage.
This is not a footnote for anyone deploying it commercially. MIT permits essentially unrestricted commercial use, sublicensing and redistribution. A custom licence may restrict any of those, may add attribution or field-of-use conditions, and may include territorial carve-outs. The obligations are whatever the text says, and the text is not interchangeable with a name.
The wider pattern is that "open source" has become a marketing register rather than a licence claim. Alibaba's Qwen models from Qwen3 onward are Apache 2.0. Google moved Gemma 4 to plain Apache 2.0. DeepSeek publishes weights under MIT with separate notes for distilled checkpoints based on Qwen and Llama. GLM-5.2 is MIT. Each of those is a different set of rights, and individual checkpoints within a family can differ from the family.
The operational rule is unglamorous and worth writing into procurement: read the licence attached to the exact checkpoint you are downloading, not the family page, not the blog post, not the comparison table. The capability gap has closed faster than the governance one.
Wavect — Best Open-Weight LLMs 2026: DeepSeek vs Qwen vs Kimi vs GLM vs Llama → · Hidekazu Konishi — Open-Weights LLM Release History and Timeline → · Hugging Face — Best Open-Source LLM Models in 2026 →