Kimi announced K2, a next-generation open-source Mixture-of-Experts (MoE) model that activates 32 billion parameters out of a 1 trillion-parameter pool.
It’s trained using a novel optimizer called MuonClip, which uses a QK-clip technique to ensure stability while maintaining token efficiency. During post-training, K2 leverages a large-scale agentic data synthesis pipeline and reinforcement learning to improve via environment interactions.
In benchmarks, it outperforms many open and closed source models in coding, mathematics, reasoning, and agentic performance. The model checkpoint is being open-sourced to further research.
.png)
.png)
