Moonshot AI's Kimi K2.7 Code cuts thinking tokens by 30% for coding
1T parameters, 32B active – open-source coding MoE lands on Hugging Face
Moonshot AI has open-sourced Kimi K2.7 Code, a Mixture-of-Experts (MoE) model designed specifically for coding and agentic tasks. With 1 trillion total parameters but only 32 billion activated per token, it strikes an efficiency-power balance that lets developers run large-scale inference cost-effectively. The model improves on its predecessor, Kimi K2.6, by reducing 'thinking tokens' by 30%—meaning it reaches answers faster without sacrificing accuracy. Benchmarks on coding and agentic tasks (e.g., code generation, tool use) show clear gains, making it competitive with other top open-source code models like DeepSeek-Coder and CodeLlama.
Kimi K2.7 Code ships under a Modified MIT License, which permits commercial use with minimal restrictions, and comes with a 256K context window capable of handling long codebases and complex prompts. By releasing on Hugging Face, Moonshot AI invites the global developer community to fine-tune, benchmark, and integrate the model into everything from IDEs to autonomous coding agents. This release signals a continued push toward open-source dominance in coding AI, challenging proprietary models with transparent, high-performance alternatives.
- 1 trillion total parameters with 32 billion activated per token, reducing compute cost
- 30% fewer 'thinking tokens' than Kimi K2.6, improving inference speed
- 256K context window under a Modified MIT License, enabling long-context coding tasks
Why It Matters
Open-source coding MoE with 1T parameters makes advanced code generation accessible for teams and startups.