Kimi K2.7-Code ships with leaner reasoning
Moonshot AI's open-weights coding model hit Hugging Face on June 12, claiming 30% fewer reasoning tokens than K2.6 with higher coding benchmark scores.
Moonshot AI published Kimi K2.7-Code to Hugging Face on June 12, an open-weights coding model released under a modified MIT license. The architecture is a 1-trillion-parameter mixture-of-experts design with roughly 32 billion parameters active per token and a 256K context window, the same scale class as its K2 predecessors.
The interesting claim is not the size. Moonshot reports the model uses about 30% fewer reasoning tokens than K2.6 while scoring higher on its coding benchmarks, an attack on the “overthinking” problem where agentic coding models burn long chains of hidden reasoning on simple edits. Moonshot’s hosted API prices it at $0.95 per million input tokens and $4.00 per million output tokens; the weights are free to self-host. The release drew a 150-plus point discussion on Hacker News within the day.
Why this matters
If you run coding agents, output tokens are your meter. A model that reaches the same answer with 30% less internal monologue is a direct cost and latency cut, which compounds across the dozens of model calls a single agent task makes. Vendor benchmark numbers deserve the usual skepticism until independent runs land, but the direction matters: open models are now optimizing for the operational economics of agents, not just leaderboard position. For what that means in practice, coding agents for non-developers covers the workflow side, and the solo operator AI stack covers when an open model beats a subscription.