Kimi K2 is an open-weight mixture-of-experts language model from Moonshot AI, with 1 trillion parameters in total, built especially for agent tasks like tool use and coding.
In plain words
In a mixture-of-experts model only some parts, the experts, run for each token, so only a small share of a huge model does the work at any moment. Kimi K2 has 1 trillion parameters but uses 32 billion per token. Its pre-training used 15.5 trillion tokens and a new optimizer called MuonClip, then taught agent skills with synthetic tasks and reinforcement learning. Base and instruct versions are released under a modified MIT licence.
A quick example
Imagine asking an agent to fix a failing test in a code project. A model like Kimi K2 would read the files, call tools to run the tests, edit the code and try again until the tests pass.
Where you'll meet it
- Coding agents and software engineering tasks.
- Chat and tool calling through an OpenAI-compatible API.
Related
Sources
- PaperKimi K2: Open Agentic Intelligence, Kimi Team (arXiv) · read 29 Sept 2026
- DocsKimi-K2, Moonshot AI (GitHub) · read 29 Sept 2026
Checked by WhatIsAI against the sources above · 29 Sept 2026. Spotted an error? Report it.