Moonshot AI open-sources Kimi K2.7 Code, a 1T-parameter MoE model for agentic coding
30% fewer reasoning tokens and state-of-the-art benchmark scores over K2.6
Moonshot AI has open-sourced Kimi K2.7 Code, a trillion-parameter mixture-of-experts (MoE) model purpose-built for agentic coding tasks. The model delivers significant benchmark improvements over its predecessor K2.6 while consuming 30% fewer reasoning tokens, making it both more capable and more efficient for deployment. Released on June 18, 2026, it lands as the second pillar of an emerging open-source coding wave, alongside GLM-5.2 from other labs.
Kimi K2.7 Code is available under open weights on Hugging Face, and Moonshot AI also launched a beta Kimi Code interface for direct experimentation. The model's 1T MoE parameters enable strong performance on code generation, debugging, and multi-step agentic workflows. This release signals a continued shift toward large, open-weights coding models that can be adapted and fine-tuned by the community.
- 1 trillion parameters in a Mixture-of-Experts (MoE) architecture
- 30% fewer reasoning tokens than K2.6 while improving benchmarks
- Released alongside GLM-5.2 as part of an open-source coding wave
Why It Matters
Open-source trillion-parameter coding models lower barriers for agentic AI development and accelerate community-driven innovation.