Moonshot AI has released Kimi-K2.7-Code, an open-source coding model designed to enhance efficiency in AI-assisted programming. The Beijing-based firm reports a 30% reduction in reasoning token usage compared to its predecessor, allowing developers to conserve compute resources while improving output quality.
The model is now available via Kimi platform APIs and on Hugging Face under a Modified MIT License, permitting commercial use with attribution. This licensing structure supports large-scale enterprise deployments.
Kimi-K2.7-Code utilizes a Mixture-of-Experts architecture featuring 1 trillion total parameters, with 32 billion active at any given time. Benchmark tests show significant performance gains: a 21.8% increase on Kimi Code Bench v2, an 11.0% rise on Program Bench, and a substantial 31.5% jump on MLS Bench Lite. The latter highlights improved multi-language support across Python, Rust, and Go.
The 30% drop in reasoning tokens mitigates "overthinking," a common issue where excessive token consumption drives up latency and API costs. This optimization targets agentic capabilities and extended context handling, enabling AI agents to plan, execute, and debug code over long sequences more effectively.
Founded in 2023 by Tsinghua University alumnus Zhilin Yang, Moonshot AI has accelerated its open-weight model releases. Following the K2 base model in July 2025, the company launched K2 Thinking, K2.5, and K2.6 in rapid succession. Kimi-K2.7-Code marks the fifth major iteration in under a year, reinforcing the company's pivot toward robust, open-source infrastructure.