The model

Moonshot AI put the weights for Kimi K2 on Hugging Face this week. It is a mixture-of-experts model with one trillion total parameters and 32 billion active, 384 experts with 8 selected per token, a 128K context and a 160K vocabulary. The pretraining run covered 15.5 trillion tokens using an optimizer they call MuonClip, and the model card claims zero training instability across the run, which if true is an achievement at this scale on its own.

The instruct model reports 89.5 percent on MMLU, 65.8 percent single-attempt on SWE-bench Verified in agentic mode, 53.7 percent on LiveCodeBench v6 and 69.6 percent on AIME 2024. Those are the numbers that got it called another DeepSeek moment and made it the most downloaded model on Hugging Face the day after release. The license is what we want to write about, because it is a small change with more in it than it looks.

The added sentence

The LICENSE file is the standard MIT text, permission to use, copy, modify, distribute and sell with the usual warranty disclaimer, plus one paragraph. If the software or any derivative is used in a commercial product or service with more than 100 million monthly active users, or more than 20 million US dollars in monthly revenue, you must prominently display Kimi K2 on the product's user interface. That is the whole modification. No use restrictions, no acceptable use policy, no separate permission needed above a threshold.

Compare that with Meta's approach, where above 700 million monthly users you must ask Meta for a license it may or may not grant. Moonshot's clause does not let them refuse anyone. It asks for credit. Below the thresholds it is pure MIT, and above them it is MIT with a logo requirement. Most companies on earth will never hit either number, so for most users this is as permissive as a model license gets.

Attribution as a lever

The design is clever in a way that copyleft used to be. The GPL used the copyright in the code to force derived works to stay open. Moonshot uses the copyright in the weights to force large derived products to say where they came from. That is a much smaller ask, but it targets the same actors: the big companies that would otherwise take an open model, build a product on it and never mention it. For a lab that is trying to establish a name outside China and raise money, a Kimi K2 label on a product with 100 million users is worth more than any royalty it could realistically collect.

We would guess this pattern spreads. DeepSeek went with plain MIT. Meta built a bespoke license. Moonshot found a middle position that costs adopters nothing until they are very large, and costs the very large adopters nothing except a credit. It is easy to imagine other Chinese labs adopting the same template with their own name in it.

Can it be enforced

Here we have to be careful, because we are a researcher and not a lawyer, and the answer depends on questions that no court has settled. The clause is a condition on a copyright license, and its force depends on whether model weights are copyrightable in the jurisdiction where the product is sold. That is an open question in the United States, where copyright requires human authorship, and it is not obviously settled elsewhere. If weights are not protected, the license is a request and the clause has no teeth.

There is a second problem, which is that derivative works of a model are hard to identify. A company that fine-tunes K2, distills it into a smaller model and ships that has arguably produced a derivative under the license. Proving it from the outside, without access to the training pipeline, is close to impossible. So in practice the clause works on companies that choose to comply, and the enforcement mechanism is reputational rather than legal.

What we would want to know

The test will come the first time a product above the threshold is built on K2 and does not display the name. Whether Moonshot sends a letter, and what happens if the recipient ignores it, will tell us whether attribution clauses are the new copyleft or a polite fiction. Until then the honest summary is that Moonshot released a trillion-parameter model under terms that are, for nearly everyone, indistinguishable from MIT, and reserved one small ask for the companies that can most easily afford it.

Sources

  1. Moonshot AI, Kimi-K2-Instruct model card
  2. Kimi-K2-Instruct LICENSE file
  3. Wikipedia, Moonshot AI