Kimi K2.7 Code

Moonshot

Open-weight model · runs on-premises · Published on 5 July 2026 · Updated on 2 September 2026

Kimi K2.7 Code delivers stability in agentic context. It maintains reliable tool calls over long sessions, supports the MCP protocol, and publishes its results on the MCP-Atlas and Program Bench agent benchmarks. Its size is close to one trillion parameters.

Key points: Agentic · Reliable tool calls · MCP protocol · MCP-Atlas · 1000B, 32B active.

Architecture

Kimi K2.7 Code is a one-trillion-parameter mixture-of-experts model with 32 billion active parameters per token (384 experts, 8 active), 61 layers, a 262,144-token context, a 400-million-parameter vision encoder and 595 GB of published weights, under the Modified MIT licence (Hugging Face card and config.json re-read on 2 September 2026). It is tuned for stability in agentic context.

Strengths

It maintains reliable tool calls over long sessions, supports the MCP protocol, and publishes its results on the MCP-Atlas and Program Bench agent benchmarks.

Use cases

Kimi K2.7 Code serves coding agents that chain many tool calls without drifting, on long code tasks.

Deployment

Served by vLLM on a GPU server, or consumed via API. It integrates with OpenCode and Hermes through the LiteLLM gateway.

Moonshot unveiled Kimi K3 in July 2026: 2,800 billion parameters, native vision, and always-on reasoning, aimed at enterprise AI. K2.7 Code remains the model of choice for coding assistants.

All models on the platform are interchangeable through a single gateway: switching models is a one-line configuration change, with no code to rewrite. The priority remains local execution.

Official sources

• moonshotai/Kimi-K2.7-Code, official model card on Hugging Face: licence, parameters, context window and weight format
• Artificial Analysis, model comparison index