Moonshot AI releases Kimi K3.1 Pro, a 2.8-trillion-parameter open-weight model

Chinese artificial intelligence company Moonshot AI released Kimi K3.1 Pro, its most capable AI model so far, on October 23, 2026. The company describes it as a "3T-class open-weight model": it has about 2.8 trillion parameters in total, of which about 104 billion are used for any one task. The model is available now through Moonshot's Kimi apps and its paid API, and the company says it will publish the full model weights for anyone to download by November 4, 2026.

A few terms explain what is new. A parameter is one of the internal numbers a model learns during training; more parameters usually mean a model can capture more knowledge and skill. A mixture-of-experts design splits the model into many specialist parts and switches on only a few for each request. That is why Kimi K3.1 Pro can have about 2.8 trillion parameters while using only about 104 billion at a time, which keeps the cost of answering a question lower. A context window is how much text the model can read at once; 1 million tokens is roughly the length of several long books.

"Open weights" means the trained model itself is published, so companies, researchers and governments can run it on their own hardware, study it and adapt it, rather than only renting access through the maker's service.

Moonshot says the model is especially good at writing code for whole projects, including games built in engines such as Unreal, Unity and Godot, and at producing 3D models in design software. These are the company's claims. On one test of AI coding agents, Terminal-Bench, Moonshot reported a score of 65.7%, close to that of Anthropic's top model. An independent ranking by the firm Artificial Analysis placed Kimi K3.1 Pro just behind the leading American model on its overall index. Such scores measure narrow tasks and do not guarantee how a model will perform in real work.

The cyber-security step is notable. As AI models get better at finding software flaws, there is a fear that publishing them openly could help attackers. Moonshot says it took cyber-related material out of parts of training for the open release, and reported that this reduced the model's score on a hacking-skills test.

For India, open-weight models matter because Indian start-ups, universities and the IndiaAI Mission can build on them without sending data to a foreign company's servers. A very large open model also raises the bar for India's own efforts to build foundation models. At the same time, Indian users and firms should note that a model's origin, licence and data rules need to be checked before it is used for sensitive work.

Short-term: Developers can use the model through Moonshot's apps and API now. Its benchmark claims will be tested by outside users over the coming days, and the release adds pressure on other labs on price.

Long-term: Once the weights are out, a model of this size becomes freely available for others to build on, which narrows the gap between open and closed AI. The approach of removing cyber data before an open release may be copied, or debated, by other labs.

Who is affected: software developers and AI start-ups, including in India; researchers who study open models; companies choosing between closed and open AI; and policymakers watching AI security.

Questions and Answers