Moonshot AI Kimi K2.7

Kimi K2.7 Code is an advanced Mixture-of-Experts coding model built for agentic software engineering, long-horizon programming, and efficient repository-scale code generation. It features 1T total parameters with 32B activated, utilising a 61-layer architecture with 7,168 hidden size and 64 attention heads. The model routes 8 experts per token across 384 routed experts and a shared expert, leveraging Multi-head Latent Attention for efficient long-context processing. It supports a 256K token context window, incorporates native INT4 quantization, and optimizes complex workflows with improved token efficiency and end-to-end task completion.