kimi-k2.7-code

Kimi K2.7-Code is a coding-focused agentic model built on top of the K2.6 generation, targeting real-world long-horizon software engineering rather than isolated code completions. The design emphasizes end-to-end task completion across complex workflows, including repository-scale understanding, multi-step refactors, and tool-driven debugging sessions. Practitioners get an open-weights release under a Modified MIT license, which lets teams self-host, fine-tune, or audit the model for proprietary codebases that require careful data handling.

The architecture follows a Mixture-of-Experts design with roughly 1T total parameters and 32B activated per token across 61 layers, with a single dense layer at the base and 384 experts selecting the active paths during inference. Moonshot reports a +21.8% gain on Kimi Code Bench v2 versus the prior K2.6 release alongside about 30% fewer thinking tokens, suggesting better quality per unit of reasoning. At launch, CoreWeave's serverless inference ranked highest on Artificial Analysis output speed and price-performance, and the model is offered with a drop-in API path, making it straightforward for teams already running earlier Kimi versions to migrate their pipelines.