1T total / 32B active
256K
Modified MIT License
Official open-weight coding-focused agent model in the Kimi K2 line, positioned by Moonshot AI as a stronger long-horizon software-engineering successor to Kimi K2.6. The model card describes Kimi K2.7 Code as a Mixture-of-Experts architecture with 1T total parameters and 32B activated parameters, 61 layers, 384 experts with 8 selected per token plus 1 shared expert, 64 attention heads, MLA attention, 160K vocabulary, and a 400M-parameter MoonViT vision encoder. Release materials emphasize end-to-end coding and agentic workflows rather than plain code completion, with claimed improvements on real-world long-horizon software-engineering tasks and about 30 percent lower thinking-token usage than Kimi K2.6. The published benchmark table highlights gains over K2.6 on Kimi Code Bench v2, Program Bench, MLS-Bench Lite, Kimi Claw 24/7 Bench, MCP-Atlas, and MCPMark-Verified, while still trailing the strongest frontier closed models on some coding evaluations. Moonshot also says the model forces thinking mode and preserve_thinking by default, retains full reasoning content across turns for coding-agent scenarios, supports interleaved thinking with multi-step tool calling, and accepts image and video inputs in addition to text. The card further notes native INT4 quantization inherited from Kimi-K2-Thinking and deployment support through vLLM, SGLang, KTransformers, and Moonshot's OpenAI/Anthropic-compatible API.