Kimi K2.7 Code is Moonshot AI's dedicated coding model, refined on top of the earlier K2.6 release to behave more like an autonomous software engineering agent. Its MoE backbone carries around one trillion total parameters with roughly 32 billion activated per token, organized across 61 layers (one dense) with 64 attention heads and a routing pool of 384 experts. That scale, paired with coding-specific post-training, lets it follow multi-step instructions inside very long sessions and finish programming tasks that span many files or services more reliably than its predecessor, while trimming unnecessary "thinking" tokens by roughly thirty percent on average.

For practitioners, the model reads as a tool for sustained, multi-turn coding workflows: long context handling, tool-calling for shell and editor actions, and structured outputs that slot into agent loops. A companion high-speed variant, kimi-k2.7-code-highspeed, exposes the same weights at about 180 tokens per second and up to 260 in short exchanges, trading a bit of stability for a more responsive interactive feel when resources are available. Weights ship under a modified MIT license on the Moonshot AI Hugging Face org, so the same checkpoint can be self-hosted or integrated alongside hosted chat surfaces, fitting teams that want a single coding-focused model for both background batch runs and live pair-programming sessions.