Kimi K2.7 Code is a coding-focused agentic model built directly on top of Kimi K2.6, sharpened for real-world, long-horizon software engineering tasks where sustained instruction following matters as much as raw code generation. According to Moonshot's documentation, the model is designed to follow instructions more reliably in long contexts and to complete multi-step coding workflows with higher success rates than its predecessor. Internal and external evaluations reported by Kimi suggest it trims roughly thirty percent of the overthinking and excess thinking tokens that K2.6 produced, which translates into more efficient reasoning traces and faster paths to working solutions on demanding programming jobs.

The architecture follows a Mixture-of-Experts design with around one trillion total parameters and about thirty-two billion activated per token, organized across sixty-one layers (including a single dense layer), sixty-four attention heads, and a pool of 384 experts that select a subset per token. The weights are openly released on Hugging Face under a Modified MIT license, making the model approachable for self-hosting, fine-tuning, and research into agentic coding pipelines. Alongside the base release, Moonshot offers a kimi-k2.7-code-highspeed variant that keeps the same underlying model but targets higher output throughput, advertising roughly 180 tokens per second and up to about 260 tokens per second in short-context scenarios, giving teams a throughput-oriented option for interactive coding experiences.