Moonshot AI ships Kimi K2.7-Code
The open-weight coding model reported a 21.8% gain over K2.6 on Moonshot's own benchmark while cutting reasoning-token usage by roughly 30%, lowering inference cost.
- Models & capabilities
- Open weights & ecosystem
- Colour
Moonshot AI released Kimi K2.7-Code, an open-weight, agentic coding model built for multi-step software engineering — planning, editing, running tools and debugging across a task — rather than general chat. On Moonshot’s own Kimi Code Bench v2, the company reported a score of 62.0 against 50.9 for its predecessor K2.6, a 21.8% gain, with smaller improvements of 9.3–31.5% across five other internal benchmarks. Moonshot also reported roughly 30% lower reasoning-token usage than K2.6, which lowers cost and latency since reasoning tokens bill as output.
The model is a mixture-of-experts design with 1 trillion total parameters and 32 billion active per token across 384 experts, a 256,000-token context window, and a vision encoder for image and video input. Weights were released on Hugging Face under a modified MIT licence.
K2.7-Code was the fifth major release in Moonshot’s Kimi series within a year, part of a rapid cadence of Chinese open-weight coding models that made this class of model a fast-moving, low-margin segment of the market; it was superseded by Kimi K3 roughly a month later.