Moonshot AI's Kimi K2.7-Code Cuts Token Usage by 30% — But There's a Catch

Moonshot AI shipped Kimi K2.7-Code on June 12, 2026, the fifth major Kimi release in under a year. The open-source model (Modified MIT license) is available on Hugging Face, via API, and the Kimi Code CLI. The headline claim is a 30% reduction in reasoning token usage compared to its predecessor, K2.6, with a 21.8% improvement on Moonshot's own Kimi Code Bench v2. However, all benchmarks are proprietary—no third-party results on SWE-bench Verified or LiveCodeBench yet. The model uses a Mixture-of-Experts design with 1 trillion total parameters but only 32 billion active per token, paired with a 256K context window. A notable design choice: the model always reasons before answering (thinking mode cannot be turned off), which affects workflow and token budgeting.

K2.7-Code shows particular strength in agentic workflows, scoring 81.1 on MCP Mark Verified for correct tool invocation via the Model Context Protocol. It also supports multimodal inputs (images and video), useful for full-stack debugging. The release is paired with Kimi Code, a terminal-first coding agent with subscriptions starting at $19/month—a platform play similar to Anthropic's Claude Code. API pricing is set at $0.95 per million input tokens and $4.00 per million output tokens. Moonshot provides migration paths for teams running K2.6 with vLLM, SGLang, or KTransformers. Analyst Mitch Ashley notes that token efficiency gains are transitory; the durable advantage lies in building governed inference constraints. Still, for DevOps teams scaling coding agents, the 30% cost reduction is immediately impactful.