Moonshot AI launched Kimi K2.7-Code, an open-source coding model that improves on its predecessor with better long-horizon task completion and 30% reduced thinking-token usage. The model targets complex software engineering workflows where token efficiency directly impacts cost and performance.

Built on a Mixture-of-Experts architecture with 1 trillion total parameters and 32 billion activated parameters, K2.7-Code features 61 layers including dense layers, 64 attention heads, and 384 experts selecting 8 per token. It supports 256K context length and uses MoonViT for vision encoding with 400M parameters.

Benchmark results show significant gains across coding and agentic tasks. On Kimi Code Bench v2, it scores 69.0 versus 67.4 for GPT-5.5 and 62.0 for Claude Opus 4.0. The model reaches 81.3 on MCP Mark Verified and 79.4 on MCP Atlas, outperforming previous versions across most evaluations. Native INT4 quantization and OpenAI-compatible API support simplify deployment.

Available through Moonshot's platform with vLLM and SGLang compatibility, K2.7-Code requires transformers version 4.57.1 or higher. The release positions Moonshot as a serious competitor in the open-source coding model space, particularly for teams needing efficient long-context processing.