Kimi K2.7-Code: open-weight 1T coding agent ships with 81.1 on MCP Mark Verified and 30% fewer reasoning tokens

Moonshot AI shipped Kimi K2.7-Code to Hugging Face on June 12 — an open-weight update to the K2.6 family, specifically tuned for long-horizon agentic software engineering.

Architecture. K2.7-Code is a 1-trillion-parameter Mixture-of-Experts model with 32 billion active parameters and 384 experts. Context window is 256K tokens. The model was trained to spend fewer tokens on "thinking": Moonshot reports roughly 30% fewer reasoning tokens than K2.6 on the same tasks, which directly reduces cost on agentic pipelines that loop over a codebase multiple times.

Benchmark claims. Moonshot reports +21.8% on Kimi Code Bench v2, +11.0% on Program Bench, and +31.5% on MLS Bench Lite compared to K2.6. The standout for MCP-native agents is 81.1 on MCP Mark Verified, which scores correct tool invocation through the Model Context Protocol. That's the metric that matters most if you're building agents that call structured external tools.

Availability and price. The model is live at the Kimi API and through Kimi Code, Moonshot's terminal-first coding agent. Pricing is $0.95 / $4.00 per million input/output tokens under the kimi-k2.7-code model ID. Weights ship under a Modified MIT license, meaning you can self-host for closed commercial products.

For teams evaluating open-weight alternatives to proprietary coding agents, the combination of sub-dollar input pricing, open licensing, strong MCP tool-call accuracy, and a 256K context window makes K2.7-Code worth benchmarking on your real task distribution before assuming a proprietary model is required.

Sources: MarkTechPost: Kimi K2.7-Code release, Crypto Briefing, llm-stats model page, Lushbinary: developer guide