Moonshot AI dropped Kimi K2.7-Code, an open-weight coding model, with the weights now live on Hugging Face under a Modified MIT license. It's the first time the Beijing company has stuck "Code" in the name. This isn't a new base model. It's a coding-tuned post-train on the existing K2 Mixture-of-Experts family, 1 trillion total parameters with 32 billion active.
The headline pitch is efficiency, not raw capability. Moonshot says K2.7 "reduces thinking-token usage by approximately 30%" against K2.6 on equivalent tasks, per the model card. Fewer tokens burned means lower cost and faster runs for agents, which is the actual selling point here.
On benchmarks the numbers are all self-reported and all measured against its own predecessor, not the frontier: +21.8% on Kimi Code Bench v2, +11.0% on Program Bench, +31.5% on MLS Bench Lite. The first of those is an in-house benchmark only Moonshot reports, so take it with the usual salt. Moonshot didn't publish SWE-Bench Verified scores against rivals at launch.
The other focus is long-horizon coding, where the model is meant to hold longer instructions and finish end-to-end tasks instead of breaking partway. A 6x high-speed mode is promised but not here yet.
It's available now through the Kimi API and the Kimi Code CLI. Pricing on the API runs around $0.95 per million input tokens and $4.00 per million output, a slight bump over K2.6.
Bottom Line
K2.7-Code is a coding post-train on the K2 MoE family that Moonshot says cuts reasoning-token usage by about 30% versus K2.6.
Quick Facts
- 1 trillion total parameters, 32 billion active (MoE)
- ~30% lower reasoning-token usage vs K2.6 (company-reported)
- +21.8% Kimi Code Bench v2, +11.0% Program Bench, +31.5% MLS Bench Lite (all self-reported)
- Released June 12, 2026 under Modified MIT license
- API pricing ~$0.95/M input, $4.00/M output tokens
- 256K context length
