The latest release of Kimi K2 (0905) marks a significant leap over the initial July edition (0711). Built on a Mixture-of-Experts architecture (1T parameters, 32B active per token), this model is not just an LLM—it’s an agentic intelligence platform optimized for programming, analysis, and large-scale projects.
What’s new in version 0905
- Extended context up to 256K tokens (double that of 0711), enabling handling of massive documents or entire code repositories.
- Enhanced coding capabilities: on SWE-Bench Verified, it rises from 65.8% to 69.2%, and on Terminal-Bench it scores +7 points over the previous release.
- Cleaner, more functional code generation: perfect for front-end and UI, producing polished and professional interfaces.
- Fewer hallucinations: maintains creativity while improving factual accuracy and long-context coherence.
- Training stability: powered by the innovative MuonClip optimizer, reducing instability in large-scale MoEs.
Superior performance:
On the GPQA-Diamond benchmark, Kimi K2-0905 leads with 75.1, outperforming closed models like GPT-4o and Claude 3.5 in mathematics and STEM reasoning.
- Pricing 2 coins per 100 words
