Kimi K2.7 Code is a coding-specialized agentic model in Moonshot AI's Kimi K2 family, built on top of Kimi K2.6 with substantial improvements for real-world long-horizon coding workflows. The architecture is a Mixture-of-Experts design with roughly 1T total parameters and 32B activated parameters, spread across 61 layers (including one dense layer), 384 experts, and 64 attention heads operating on a 7168-dimension attention hidden state. It accepts text and image inputs natively, outputs text, and always runs in a thinking mode that preserves full reasoning content across multi-turn conversations, which suits it to step-by-step programming and agentic task decomposition.

Designed for end-to-end software engineering work, the model targets long-context programming with a 256K-token window and emphasizes token efficiency, cutting thinking-token usage by about 30% compared with its predecessor Kimi K2.6. It supports structured outputs, tool calling, and attachment handling, making it a practical fit for complex pipelines where the model must read code, plan changes, invoke external tools, and return well-formed results. Open weights are available, so teams can self-host for reproducible debugging and integration into internal coding assistants that need sustained reasoning over large codebases.