KIMI K2

Open Agentic Intelligence

Technical Report TLDR

Executive Summary

KIMI K2 is a 1 trillion-parameter MoE LLM with 32B activated parameters. It introduces the MuonClip optimizer for stable training and achieves SOTA performance among open-source non-thinking models. Key achievements include 65.8 on SWE-Bench Verified, 47.3 on SWE-Bench Multilingual, and 76.5 on ACEBench (En) - all without extended thinking.

SWE-Bench Verified
65.8
vs 54.6 (DeepSeek-V3)
SWE-Bench Multilingual
47.3
vs 31.5 (DeepSeek-V3)
LiveCodeBench v6
53.7
SOTA for non-thinking models
ACEBench (En)
76.5
Agentic capabilities
AIME 2025
49.5
Mathematical reasoning
GPQA-Diamond
75.1
STEM performance

Performance Comparison: Key Benchmarks

Agentic Intelligence Revolution

KIMI K2 represents a paradigm shift from static imitation learning to active learning through interactions. This enables superhuman capabilities through autonomous exploration and exploitation.

Key Innovations:

  • MuonClip Optimizer: Novel QK-clip technique for stable training
  • 15.5T Tokens: Zero loss spike pre-training
  • Multi-stage RL: Large-scale agentic data synthesis

Technical Achievements

K2 demonstrates exceptional performance across software engineering, coding, mathematics, and agentic tasks without extended thinking - positioning it as one of the most capable open-source LLMs to date.