Kimi K2.7 Code: Open-Source 1T Model Beats Claude on Tool Use

Moonshot AI released Kimi K2.7 Code on June 12, 2026, the latest in the Chinese AI lab's rapidly iterating Kimi K2 series. The model features a trillion-parameter Mixture-of-Experts architecture with 32 billion active parameters across 384 experts (8 routed plus 1 shared), a 256K context window, and native INT4 quantization for efficient inference.

The headline benchmark result: Kimi K2.7 Code scores 81.1% on MCPMark — a tool-use and agentic coding benchmark — compared to Claude Opus 4.8's 76.4%. It achieves this while using 30% fewer thinking tokens than its predecessor, K2.6. The model also includes MoonViT, a ~400M parameter vision encoder for multimodal inputs.

What makes this release significant is the combination of open weights and aggressive pricing. The model is available on Hugging Face under a Modified MIT license, deployable via vLLM, SGLang, and KTransformers, and priced at $0.95 per million tokens via Moonshot's API. Moonshot also launched Kimi Code CLI, a coding assistant starting at $19/month that competes directly with Claude Code and GitHub Copilot.

For developers evaluating coding models in 2026, the landscape has shifted dramatically. Open-weight models from Chinese labs are no longer a step behind — in agentic tool use, they're leading. The Modified MIT license requires attribution but imposes no commercial restrictions, making Kimi K2.7 Code viable for enterprise deployment. Builders who want sovereignty over their inference stack now have a genuine open-weight alternative to Anthropic and OpenAI for coding agents.