Moonshot AI has released Kimi-K2.7-Code, an open-source model designed for software development and agentic coding workflows. The company says the new system builds on its earlier Kimi K2.6 model and is tuned for longer, more complex programming tasks that require sustained execution across multiple steps.
According to Moonshot AI, Kimi-K2.7-Code is intended to improve end-to-end completion on real-world coding jobs, not just short code-generation prompts. The model is positioned as a coding-focused agent that can handle broader software engineering workflows, including tasks that involve planning, implementation and iterative refinement. Moonshot says it also uses fewer thinking tokens than Kimi K2.6, with roughly a 30 percent reduction in thinking-token usage.
Moonshot describes Kimi-K2.7-Code as a mixture-of-experts model with 1 trillion total parameters and 32 billion activated parameters. The architecture includes 61 layers, 384 experts, and a context length of 256,000 tokens. The model also uses native int4 quantization, a technique Moonshot says it shares with Kimi-K2-Thinking. The company lists support for image and video input as part of the model’s usage examples.
The release notes say Kimi-K2.7-Code forces thinking mode and keeps preserve_thinking enabled by default. That setup is meant to retain reasoning across multiple turns, which Moonshot says improves performance in coding-agent scenarios. The model is also presented as compatible with OpenAI- and Anthropic-style APIs through Moonshot’s platform.
Moonshot says the model can be deployed with inference engines including vLLM, SGLang and KTransformers. It adds that deployment methods used for Kimi-K2.5 and Kimi-K2.6 can be reused because the architecture is the same. The company lists a Transformers version requirement of 4.57.1 or later, but lower than version 5.
Moonshot shared benchmark results that compare Kimi-K2.7-Code with Kimi K2.6, GPT-5.5 and Claude Opus 4.8 across coding and agentic tasks. In the company’s table, Kimi-K2.7-Code scored higher than Kimi K2.6 on each benchmark listed.
On Kimi Code Bench V2, Moonshot reported a score of 69.0 for Kimi-K2.7-Code, up from 62.0 for Kimi K2.6. On Program Bench, the new model scored 69.1 versus 53.6 for Kimi K2.6. On MLS Bench Lite, it scored 35.5, compared with 26.7 for the earlier model. In the agentic category, Moonshot listed 52.8 on its Kimi Claw 24/7 Bench, 79.4 on MCP Atlas, and 92.9 on MCP Mark Verified.
The company’s benchmark notes say the tests were run under specified configurations, including thinking mode for Kimi models and different high-effort settings for competing systems. Moonshot also says some of the benchmarks, such as Kimi Claw 24/7 Bench and MCPMark-Verified, are in-house or human-verified evaluations.
Moonshot is making the model available through its platform at platform.moonshot.ai and on Hugging Face. The release adds another open-source option in the growing market for coding models aimed at agentic software development and long-context tasks.