Moonshot releases Kimi K2, an open agentic MoE model with 32B activated parameters

Moonshot has released Kimi K2, a Mixture-of-Experts large language model featuring 1 trillion total parameters and 32 billion activated parameters. The model is designed for agentic intelligence, utilizing a novel MuonClip optimizer to pre-train on 15.5 trillion tokens without loss spikes.

Kimi K2 achieves state-of-the-art performance among open-source non-thinking models, scoring 66.1 on Tau2-Bench, 76.5 on ACEBench (En), and 65.8 on SWE-Bench Verified. It also demonstrates strong capabilities in coding and reasoning, with scores of 53.7 on LiveCodeBench v6 and 49.5 on AIME 2025.

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