Moonshot AI launches Kimi-K2.7-Code with 1T parameters, cuts reasoning tokens by 30%
Open-source coding model slashes compute waste while boosting benchmarks across Python, Rust, Go.
Moonshot AI, the Beijing-based startup behind the Kimi chatbot, has open-sourced Kimi-K2.7-Code, a 1 trillion parameter coding model built on a Mixture-of-Experts architecture with only 32 billion active parameters. The model delivers a 30% reduction in reasoning tokens compared to its predecessor, directly tackling the overthinking problem common in AI-assisted coding—where models waste compute and increase latency by over-analyzing. Benchmark gains are substantial: 21.8% on Kimi Code Bench v2, 11.0% on Program Bench, and 31.5% on MLS Bench Lite (multilanguage support across Python, Rust, Go). The model is released under a Modified MIT License that permits commercial use with attribution for large-scale deployments, and is accessible via Hugging Face and the Kimi API.
Founded in 2023 by Zhilin Yang, Moonshot AI has rapidly iterated its K2 series: K2 base (July 2025), K2 Thinking (Nov 2025), K2.5 (Jan 2026), K2.6 (Apr 2026), and now K2.7-Code (June 2026). The model emphasizes agentic capabilities and extended context handling, making it well-suited for planning, executing, and debugging long sequences of code. By reducing compute costs while improving accuracy across languages, Kimi-K2.7-Code positions itself as a practical tool for developers building AI-powered coding assistants or automating complex software tasks.
- 30% reduction in reasoning token usage reduces compute overhead and API costs for developers.
- Benchmarks show up to 31.5% improvement (MLS Bench Lite) for multilanguage coding tasks across Python, Rust, Go.
- Modified MIT License permits commercial use with attribution for large-scale deployments.
Why It Matters
Open-source coding model cuts waste and boosts accuracy, making AI-assisted programming more efficient for enterprise development.