Kimi K2.7 Code is Moonshot AI's coding-focused agentic refinement of the Kimi K2.6 base, shaped for end-to-end work on complex software engineering tasks. The model card frames it as a long-horizon coding assistant designed to carry multi-step development workflows through to completion, while cutting down on the thinking tokens it burns along the way by roughly thirty percent compared with its predecessor. That combination of agentic behavior and tighter token efficiency points to a model tuned less for one-shot answers and more for sustained, tool-driven development sessions where planning, editing, and verification happen together.

The model sits on a Mixture-of-Experts backbone with one trillion total parameters and about thirty-two billion activated per token across sixty-one layers, including a single dense layer. Each token is routed through a selection of experts drawn from a pool of three hundred eighty-four, with an attention hidden dimension of 7168 and sixty-four attention heads, giving it the capacity headroom for long-context coding while keeping inference cost closer to that of a much smaller dense model. Distributed under a Modified MIT license on the public Moonshot AI Hugging Face organization, it fits developers who want a self-hostable, agentic coding model that can be folded into existing pipelines without proprietary licensing constraints.