Moonshot AI has released the Kimi-K2-Instruct model card and checkpoints for its Kimi K2 mixture-of-experts language model. The model features 32 billion activated parameters out of a total of 1 trillion, trained on 15.5T tokens using the Muon optimizer.
- Kimi K2 is designed for frontier knowledge, reasoning, and coding tasks with specific optimization for agentic capabilities.
- It achieves 65.8% pass@1 on SWE-bench Verified tests with bash/editor tools and 47.3% on SWE-bench Multilingual tests.
- The model supports tool calling, autonomous problem-solving, and messages with name fields.
- Checkpoints are available in block-fp8 format on Hugging Face for use with vLLM, SGLang, KTransformers, and TensorRT-LLM.
The release provides both a base model for fine-tuning and the post-trained Kimi-K2-Instruct variant for general-purpose chat and agentic experiences.