Kimi K2.7 Code: 1T-Parameter Open-Source Coding Model Released
Moonshot AI quietly dropped one of the most capable open-weight coding models on June 12, 2026. Kimi K2.7 Code is a 1-trillion-parameter Mixture-of-Experts (MoE) model with 32B active parameters across 384 experts (8 selected per token), built on the Kimi K2.6 architecture with 61 layers.
The model is specifically tuned for agentic coding workflows — long-horizon software engineering tasks, tool invocation, multi-file refactoring, and repository-scale reasoning. Moonshot reports a 21.8% improvement on their internal Kimi Code Bench v2 and 30% fewer reasoning tokens compared to K2.6, meaning it's both smarter and cheaper to run.
Key specs: 262K context window, MoonViT vision encoder (400M parameters) for multimodal input, native INT4 quantization for efficient deployment, and compatibility with vLLM and SGLang inference engines. It's available on Hugging Face under Modified MIT and via the Kimi API platform.
Moonshot also launched Kimi Code (KFC) — a CLI-based developer toolkit with Turbo-speed models and subscription plans starting at $19/month, positioning K2.7 Code as the backend engine for a full agentic coding workflow.
The open-weight release means any team can self-host a frontier coding model. For regions with data sovereignty requirements — including much of MENA — this is significant: you don't need to send proprietary code to a US API.
Self-hostable frontier coding models are the most underrated trend of 2026. For MENA builders operating under data residency constraints or with sensitive codebases, Kimi K2.7 Code removes the 'send everything to OpenAI' dependency. The 32B active parameter footprint means it runs on a single H100 node — real infrastructure, but accessible.
Can Kimi K2.7 Code be self-hosted?
Yes. The model weights are available on Hugging Face under Modified MIT license. With native INT4 quantization and vLLM/SGLang support, it can be deployed on a single H100 node with sufficient VRAM.