Moonshot AI Kimi K2.7 Code now available on Workers AI

Moonshot AI Kimi K2.7 Code is now available on Workers AI under the identifier cf/moonshotai/kimi-k2.7-code. The model is a code-optimized variant of the Kimi K2 family. It uses a Mixture-of-Experts architecture with 1 trillion total parameters and 32 billion active parameters per token. Kimi K2.7 Code records gains compared to Kimi K2.6 on coding and agentic benchmarks. Performance improves by 21.8 percent on Kimi Code Bench v2, by 11.0 percent on Program Bench, and by 31.5 percent on MLS Bench Lite. The model consumes 30 percent fewer reasoning tokens than the previous version. The model supports a context window of 262.1k tokens. This capacity allows retention of full conversation history, tool definitions, and codebases during extended agent sessions. It also accepts vision inputs for image and text processing together. Additional capabilities include thinking mode with adjustable reasoning depth through chat_template_kwargs.thinking. Multi-turn tool calling enables agents to use tools over several conversation turns. Structured outputs are available with JSON schema support. For users moving from Kimi K2.6, the new model focuses on coding tasks with improved benchmark results and reasoning efficiency. Cached input token pricing stands at 0.19 dollars per million tokens, compared to 0.16 dollars for K2.6. API usage stays the same and requires no parameter adjustments. Access occurs via the Workers AI binding with env.AI.run, the REST API endpoint at /ai/run, or the OpenAI-compatible path at /v1/chat/completions. AI Gateway works with these endpoints as well. Further details appear on the Kimi K2.7 Code model page and pricing documentation.