From the archive
Moonshot releases Kimi K2 for coding and tool use
Kimi K2 arrives with base and instruction-tuned weights, hosted access, and support for tool-calling applications.
Moonshot AI released Kimi K2, a mixture-of-experts language model designed for coding and tasks that involve external tools. The model has one trillion total parameters, with 32 billion activated for each token.
Two downloadable variants
The release provides two starting points:
- Kimi-K2-Base for further training and custom work.
- Kimi-K2-Instruct for chat and tool-calling applications.
The original model uses a 128,000-token context window. The model weights and code are released under a modified MIT license, so the exact release terms matter when planning redistribution or commercial deployment.
Tools need an application around them
Kimi K2 can request a function call, but an application supplies and executes the actual tools. That surrounding code controls which functions are available and what permissions they have. Coding and data-analysis workflows therefore depend on both the model and the execution environment.
Moonshot offered hosted access alongside the downloadable weights. For self-managed serving, its deployment guidance includes vLLM . The Moonshot / Kimi profile covers current model access and developer options.