Moonshot AI Releases Kimi K2.7 Code: A 1T Parameter MoE Model for Advanced Software Engineering
Moonshot AI has introduced Kimi K2.7 Code, an open-weight, coding-centric Large Language Model featuring a massive 1 trillion parameter Mixture-of-Experts (MoE) architecture, designed to compete with top-tier proprietary models in tool-calling and software development tasks.
Architectural Specifications
Kimi K2.7 Code leverages a Mixture-of-Experts (MoE) architecture to balance massive capacity with computational efficiency. The model boasts a total of 1 trillion parameters, with 384 experts, of which only 32 billion are active during any single inference pass. This design allows the model to maintain a vast internal knowledge base while keeping latency and compute costs manageable.
To support complex software engineering projects and large-scale codebase analysis, the model features a substantial context window of 256K tokens, enabling it to process extensive documentation and multiple source files simultaneously.
Performance and Benchmarks
According to vendor-reported data, Kimi K2.7 Code demonstrates particular strength in tool-calling capabilities. On the MCP Mark benchmark, the model achieved a score of 81.1%, surpassing Claude Opus 4.8, which scored 76.4%.
In general coding proficiency measured by the Kimi Code Bench v2, the model scored 62.0. While this trails behind GPT-5.5 (69.0) and Claude Opus (67.4), it establishes the model as a highly competitive open-weight alternative for automated coding tasks.
Deployment, Licensing, and Integration
Moonshot AI has released the model weights on