Kimi K2 is a series of large language models developed by the Moonshot AI team, adopting a Mixture-of-Experts (MoE) architecture with 32 billion activated parameters and 1 trillion total parameters.
In terms of main capabilities, official materials show that it used the Muon optimizer during training and applied it at an unprecedented scale, while also developing new optimization techniques to address instability issues during scaling; pretraining was completed on 15.5T tokens, with zero instability during the training process. The model performs excellently on frontier knowledge, reasoning, and coding tasks, and was specifically designed for agent capabilities such as tool use, reasoning, and autonomous problem solving.
The model series includes two versions: Kimi-K2-Base is the base model, suitable for researchers and developers who want full control over fine-tuning and customization plans; Kimi-K2-Instruct is the post-trained model, suitable for plug-and-play general chat and agent experiences, and is a reactive-level model that does not perform long thinking.
In terms of usage, it can be accessed through https://platform.moonshot.ai to access the Kimi K2 API, and the official offering provides APIs compatible with OpenAI/Anthropic. Among them, the Anthropic-compatible API maps temperature through real_temperature = request_temperature * 0.6, to better be compatible with existing applications. Model checkpoints are stored in block-fp8 format and can be obtained on Huggingface. Currently, it is recommended to run Kimi-K2 on inference engines such as vLLM and SGLang, and deployment examples can be found in the Model Deployment Guide. After the local inference service is set up, interaction can be carried out through the chat endpoint; the recommended temperature for Kimi-K2-Instruct is temperature = 0.6, and if there are no special requirements, the system prompt is a good default choice.
Kimi-K2-Instruct has strong tool-calling capabilities and requires passing in the list of available tools in each request, and the model will autonomously decide when and how to call them. The official example provides an end-to-end example of calling a weather tool, in which the tool_call_with_client function implements the process from user query to tool execution, and this process requires the inference engine to support Kimi-K2's native tool parsing logic; for streaming output and manual tool parsing, see the Tool Calling Guide.
Regarding licensing and contact, both the code and model weights are released under the Modified MIT License. If you have questions or concerns, you can contact support@moonshot.cn.
