Chinese AI startup Moonshot AI has released Kimi K2.7-Code, the latest update to its K2 coding model family, and the headline number is efficiency: the company claims the new model uses roughly 30% fewer reasoning tokens than its predecessor, K2.6.
For anyone who has watched AI inference costs climb alongside capability, that claim matters. Reasoning tokens — the internal "thinking" steps some models perform before giving an answer — are expensive. Cutting them by nearly a third, if the claim holds up in independent testing, could make the model meaningfully cheaper to run at scale.
According to MarkTechPost, Kimi K2.7-Code is a coding-focused, agentic model built on top of K2.6, featuring a 256,000-token context window — large enough to hold tens of thousands of lines of code in a single session. Moonshot reports performance gains over K2.6 across six benchmarks, with the most notable being a +21.8% improvement on the company's own Kimi Code Bench v2.
As Sean Michael Kerner reported in VentureBeat, the model is available under a modified MIT license, placing it in the growing category of open-weight models that developers can download, inspect, and adapt — a meaningful contrast to the closed systems offered by OpenAI and Anthropic.
The agentic framing is notable: Moonshot is positioning K2.7-Code not just as a code completer but as a model capable of taking multi-step actions, which aligns with the industry's current push toward AI that can autonomously handle longer software development tasks.
As AI coding tools become central to how software gets built, a more token-efficient reasoning model could lower the cost barrier for startups and individual developers who want powerful code assistance without a steep API bill.