Moonshot, the developer behind the Kimi K2 model, has released a detailed technical report outlining the advancements and methodologies employed in the model's development. The report highlights the introduction of the MuonClip optimizer, a large-scale agentic data synthesis pipeline that generates tool-use demonstrations through both simulated and real-world environments, and a reinforcement learning (RL) framework that integrates RLVR with a self-critique rubric reward mechanism. The MuonClip optimizer demonstrated stable training after 70,000 iterations, with the QK-clip component becoming inactive without any loss in performance, a notable achievement at smaller scales. The Kimi K2 model is recognized as a leading non-reasoning model, building upon prior iterations such as Kimi 1.5, which featured innovative RL approaches predating improvements like Dr. GRPO. Additionally, Kimi K2 has been successfully deployed as a transformers.js application on Hugging Face, facilitating easier integration and use. The technical report has been well received within the AI community, emphasizing the model's design and performance improvements.