Researchers introduce Kimi K2, a Mixture-of-Experts large language model featuring 32 billion activated parameters out of a total of one trillion. The model utilizes the novel MuonClip optimizer to address training instability while maintaining token efficiency during pre-training on 15.5 trillion tokens.

  • Kimi K2 achieves state-of-the-art performance among open-source non-thinking models, scoring 66.1 on Tau2-Bench and 76.5 on ACEBench (En).
  • It reaches 65.8 on SWE-Bench Verified and 47.3 on SWE-Bench Multilingual, surpassing most open and closed-sourced baselines in non-thinking settings.
  • The model demonstrates strong capabilities in coding and reasoning with scores of 53.7 on LiveCodeBench v6, 49.5 on AIME 2025, and 75.1 on GPQA-Diamond.
  • Post-training involves a multi-stage process including large-scale agentic data synthesis and joint reinforcement learning to improve agentic capabilities.

The release of base and post-trained model checkpoints aims to facilitate future research and applications of agentic intelligence.