Kimi K2.7 Code is a coding-focused agentic model built on top of Kimi K2.6, designed to handle real-world long-horizon software engineering work rather than short, isolated snippets. It is published by Moonshot AI with open weights under a Modified MIT license, making the full model available on Hugging Face and deployable through the Fireworks AI model path accounts/fireworks/models/kimi-k2p7-code. The model's emphasis on end-to-end task completion positions it for agentic workflows where a system needs to plan, edit, run, and verify across multiple steps of a project.

Compared with its predecessor, the release claims substantial improvements on long-horizon coding benchmarks while reducing thinking-token usage by roughly 30%, yielding better token efficiency for extended agentic runs. The Hugging Face model card documents a Mixture-of-Experts architecture with one trillion total parameters, 32 billion activated per token, 61 layers (including one dense layer), 64 attention heads, 384 experts, and a 7168-dimension attention hidden size with 2048 per expert, reflecting a sparse design tuned for code reasoning at scale. Practically, the model is exposed through Kimi Code chat, a playground, serverless API access, on-demand deployment, and LoRA-based fine-tuning, giving teams multiple ways to integrate it into development pipelines.