Moonshot AI's latest coding model slashes reasoning token usage by 30% while posting double-digit benchmark gains over its predecessor
Moonshot AI just dropped Kimi-K2.7-Code, an open-source coding model that wants to make AI-assisted programming less wasteful and more capable. The Beijing-based company claims the model cuts reasoning token usage by 30% compared to its predecessor, which in practical terms means developers burn through fewer compute resources while getting better results.
The model is live on Moonshot AI’s Kimi platform APIs and hosted on Hugging Face under a Modified MIT License. That license permits commercial use with attribution for large-scale deployments, a detail that matters for any company thinking about building products on top of it.
The numbers behind the upgrade #
Kimi-K2.7-Code is a Mixture-of-Experts architecture packing 1 trillion total parameters with 32 billion active parameters.
The benchmark improvements over the previous K2.6 model are hard to ignore. Moonshot AI reports a 21.8% gain on Kimi Code Bench v2, an 11.0% improvement on Program Bench, and a 31.5% jump on MLS Bench Lite.
That last number is particularly striking. MLS Bench Lite tests multi-language support capabilities, meaning the model handles tasks across programming languages like Python, Rust, and Go with meaningfully better accuracy than before.
The 30% reduction in reasoning tokens addresses what researchers call “overthinking,” a common problem in automated coding environments. When an AI model spends too many tokens reasoning through a problem, it burns compute, increases latency, and drives up API costs for developers.
From chatbot startup to open-source powerhouse #
Moonshot AI was founded in 2023 by Zhilin Yang, a Tsinghua University alumnus who built the company around its Kimi chatbot. The pivot toward open-weight model releases started with the K2 series in mid-2025, and the pace of iteration since then has been relentless.
The K2 base model launched in July 2025. K2 Thinking followed in November 2025, adding enhanced reasoning capabilities. K2.5 arrived in January 2026, and K2.6 came in April 2026. Now K2.7-Code lands in June 2026, making it the fifth major release in under a year.
The company has positioned its models around three pillars: agentic capabilities, extended context handling, and multimodal inputs. K2.7-Code leans heavily into the first two, designed for scenarios where an AI agent needs to plan, execute, and debug code across long sequences.
Disclosure: This article was edited by Editorial Team. For more information on how we create and review content, see our