We introduce Kimi K2, a Mixture-of-Experts (MoE) large language model with 32 billion activated parameters and 1 trillion total parameters. We propose the MuonClip optimizer, which improves upon Muon with a novel QK-clip technique to address training instability while enjoying the advanced token efficiency of Muon. Based on MuonClip, K2 was pre-trained on 15.5 trillion tokens with zero loss spike. During post-training, K2 undergoes a multi-stage post-training process, highlighted by a large-scale agentic data synthesis pipeline and a joint reinforcement learning (RL) stage, where the model improves its capabilities through interactions with real and synthetic environments. Kimi K2 achieves state-of-the-art performance among open-source non-thinking models, with strengths in agentic capabilities. Notably, K2 obtains 66.1 on Tau2-Bench, 76.5 on ACEBench (En), 65.8 on SWE-Bench Verified, and 47.3 on SWE-Bench Multilingual -- surpassing most open and closed-sourced baselines in non-thinking settings. It also exhibits strong capabilities in coding, mathematics, and reasoning tasks, with a score of 53.7 on LiveCodeBench v6, 49.5 on AIME 2025, 75.1 on GPQA-Diamond, and 27.1 on OJBench, all without extended thinking. These results position Kimi K2 as one of the most capable open-source large language models to date, particularly in software engineering and agentic tasks. We release our base and post-trained model checkpoints to facilitate future research and applications of agentic intelligence.
核心贡献 · Key contributions
提出 MuonClip 优化器，集成 QK-Clip 技术，实现 1T MoE 模型在 15.5T token 上的稳定预训练。 Proposes MuonClip optimizer integrating QK-Clip for stable pre-training of 1T MoE model on 15.5T tokens.
开发大规模智能体数据合成流水线，通过模拟和真实环境生成多样化的工具使用轨迹。 Develops large-scale agentic data synthesis pipeline generating diverse tool-use trajectories via simulated and real environments.
设计联合强化学习框架，结合可验证奖励和自我批评准则奖励，用于后训练对齐。 Designs joint RL framework combining verifiable rewards and self-critique rubric reward for post-training alignment.
在 Tau2-Bench、ACEBench、SWE-Bench 上取得最先进的智能体性能，超越开源和闭源模型。 Achieves state-of-the-art agentic performance on Tau2-Bench, ACEBench, SWE-Bench, surpassing open and closed models.
在 LiveCodeBench、AIME、GPQA-Diamond 上展现强大的编码、数学和推理能力，取得最高分。 Demonstrates strong coding, math, and reasoning capabilities with top scores on LiveCodeBench, AIME, GPQA-Diamond.
开源基座和后训练检查点，以促进智能体智能的研究和应用。 Open-sources base and post-trained checkpoints to facilitate research and applications of agentic intelligence.
局限 · Limitations
在困难推理任务上，性能可能因过度生成 token 或不完整工具调用而下降。 Performance on hard reasoning tasks may degrade with excessive token generation or incomplete tool calls.
完整软件项目的一次性提示成功率低于在智能体编码框架下使用 K2。 One-shot prompting success rate for complete software projects is lower than using K2 under agentic coding framework.
合成数据生成可能引入幻觉或意外毒性，需要仔细的保真度验证。 Synthetic data generation may introduce hallucinations or unintended toxicity, requiring careful fidelity verification.
智能体数据合成中的模拟保真度限制可能影响现实世界的泛化能力。 Simulation fidelity limitations in agentic data synthesis may affect real-world generalization.
自动红队评估存在主观性，且对无工具调用能力的基座 LLM 相关性有限。 Automated red-teaming evaluation involves subjectivity and limited relevance for base LLMs without tool-calling.
论文章节 · Sections（共 18）
摘要Abstract
1 引言1 Introduction
2 预训练2 Pre-training
2.1 MuonClip：基于权重裁剪的稳定训练2.1 MuonClip: Stable Training with Weight Clipping
2.2 预训练数据：通过改写提升词元利用率2.2 Pre-training Data: Improving Token Utility with Rephrasing