Kimi K2.7 Code is a coding-focused large language model developed by Moonshot AI, specifically engineered to complete end-to-end software engineering tasks reliably over long contexts. Built on the Kimi K2 architecture, the model is optimized for long-horizon coding, agentic task decomposition, and complex debugging sessions. It is a native multimodal model capable of processing text, images, and video, allowing it to interpret visual contexts such as UI designs, terminal screenshots, and demonstration videos within a single development workflow.
The model utilizes a native Mixture-of-Experts (MoE) architecture with 1 trillion total parameters, activating approximately 32 billion parameters per token. Its internal structure consists of 384 experts, including one shared expert, across 61 layers. To handle long-context information efficiently, Kimi K2.7 Code implements Multi-head Latent Attention (MLA). Compared to its predecessor, Kimi K2.6, this version reduces reasoning token usage by approximately 30% while maintaining or exceeding performance across major coding benchmarks, such as Kimi Code Bench v2.
A central feature of Kimi K2.7 Code is its mandatory "Thinking" mode, which forces the model to preserve full reasoning content across multi-turn interactions. This design prioritizes planning and verification for engineering tasks, ensuring that logical steps are transparent and consistent throughout a session. The model supports a context window of 262,144 tokens, enabling it to ingest large codebases and extensive technical documentation. Official guidance suggests using a temperature of 1.0 and a top_p of 0.95 when the thinking mode is active to achieve optimal results.
Kimi K2.7 Code is released with open weights under a Modified MIT License, which permits commercial use with attribution. It is highly optimized for agentic workloads, supporting multi-turn tool calling, structured JSON output, and automatic context caching. These capabilities make it suitable for integration into autonomous software agents and complex developer toolchains that require sustained reasoning over large-scale software projects.