Moonshot launches Kimi K3.1 Pro, promises open weights

Moonshot AI, the Chinese company behind the Kimi assistant, unveiled Kimi K3.1 Pro on Friday. The model is already running in Kimi.com, the mobile apps and the API, but it can't be downloaded yet.

The weights, the files that let anyone download a model and run it on their own servers, are due "by Nov. 4." Until then the model is open only as a promise, and the announcement doesn't say what license it will carry.

K3.1 Pro is also in Kimi Work, Kimi Code and the API, listed as `kimi-k3.1-pro` and `kimi-k3.1-pro-highspeed`. Moonshot says the second variant runs two to three times faster. In Kimi Business it ships switched off until an administrator turns it on. The announcement gives no pricing and says nothing about availability in Mexico.

The model is Kimi K3 with additional reinforcement learning, the training step that rewards correct answers. Moonshot puts it in the 3-trillion-parameter class, with a context window of 1 million tokens, the amount of text it can read at once.

All performance figures come from the company, and none has been independently verified. On Terminal-Bench 4.0, a test of coding agents, the model scores 65.7%. GPT-6 Astra gets 58.2% and Claude Opus 5.5 gets 66.4%. On its toughest hallucination test, false claims fall from 23.4% for K3 to 8.3%. Moonshot adds that the test measures an extreme case, not everyday error rates.

The unusual part is what Moonshot took out. It says it removed data tied to cyberattacks from both training stages, so as not to release an open model that raises the ceiling on cyber capability. On an in-house exploit test, the released version solves 81.0% of problems in three tries; a checkpoint without those measures reached 90.5%.

For a development team in Mexico, the payoff would be practical: with the weights in hand, the model can run on in-house infrastructure and the data never leaves the company, which the paid API can't offer. All of it hinges on Moonshot keeping its Nov. 4 date and on the license.