Moonshot AI has released Kimi K2.7 Code on Hugging Face, adding a new coding-focused model to its Kimi line that the company says is designed for agentic software work. The model is built on Kimi K2.6 and is aimed at handling longer, more complex programming tasks from start to finish.
According to the model card, Kimi K2.7 Code is intended to improve end-to-end completion across real-world software engineering workflows. Moonshot AI says the new version performs better on long-horizon coding tasks and uses fewer thinking tokens, cutting that usage by about 30% compared with Kimi K2.6. The company presents the release as an upgrade for coding agents that need to plan, reason and execute over extended task chains.
Kimi K2.7 Code uses a mixture-of-experts architecture. The model card lists 1 trillion total parameters, with 32 billion activated parameters per token. It has 61 layers, 384 experts and selects 8 experts per token. Moonshot AI also says the model supports a context length of 256,000 tokens, which positions it for very large codebases and long multi-step interactions.
The release includes a 160,000-token vocabulary, an MLA attention mechanism, and a MoonViT vision encoder with 400 million parameters. Moonshot AI says the model can take image and video inputs, although it notes that some multimodal features are only supported in its official API for now.
Moonshot AI published results across coding and agentic benchmarks, comparing Kimi K2.7 Code with Kimi K2.6, GPT-5.5 and Claude Opus 4.8. On the company’s in-house Kimi Code Bench V2, Kimi K2.7 Code scored 69.0, ahead of Kimi K2.6 at 62.0, GPT-5.5 at 67.4 and Claude Opus 4.8 at 67.4.
The model also led on Program Bench in the figures provided, scoring 69.1 versus 48.3 for Kimi K2.6, 53.6 for GPT-5.5 and 63.8 for Claude Opus 4.8. On MLS Bench Lite, Kimi K2.7 Code scored 35.5, slightly above Kimi K2.6 at 26.7 and GPT-5.5 at 35.1, while Claude Opus 4.8 scored 42.8.
For agentic performance, Moonshot AI highlighted results on its Kimi Claw 24/7 Bench and on MCP-oriented tool-use tests. Kimi K2.7 Code scored 52.8 on the 24/7 benchmark, compared with 42.9 for Kimi K2.6, 46.9 for GPT-5.5 and 50.4 for Claude Opus 4.8. On MCP Atlas, it scored 79.4, and on MCPMark-Verified it scored 92.9.
The company says the evaluations were run under specified settings, including thinking mode for Kimi models and standardized tool-use budgets for the MCP tests. It also notes that the benchmark comparisons were averaged over multiple runs.
Moonshot AI says the model is available through its API platform, with OpenAI- and Anthropic-compatible interfaces. It recommends deployment with vLLM, SGLang or KTransformers, and says the architecture is compatible with the same deployment method used for prior Kimi releases.
The company lists a Transformers version requirement of at least 4.57.1 and below 5.0.0. The Hugging Face page also points users to deployment guidance and example code for chat, image and video input.
Kimi K2.7 Code is released under a modified MIT license. Moonshot AI says the model forces thinking mode and preserves reasoning across multi-turn interactions, a design choice it says is meant to improve performance in coding agent scenarios.