Kimi (Moonshot AI) has released and open-sourced its latest programming model, Kimi-K2.7-Code, and said it delivers benchmark gains and lower inference token usage compared with K2.6. According to Foresight News, the company reported a 21.8% improvement on Kimi Code Bench v2, an 11.0% increase on Program Bench, and a 31.5% rise on MLS Bench Lite.
Kimi said inference token usage was reduced by 30%, alongside improved instruction-following for long-chain programming tasks and higher end-to-end success rates.
The model is available via the Kimi API and Kimi Code, and the company said a six-times speed mode will be introduced soon.
On the same day, Kimi launched the Kimi Code Beta program, allowing users to apply for early access to unreleased models and features.