moonshotai/Kimi-K2.7-Code

Hugging Face
VISIONConcurrency Cost:4Model Size:1000BQuant:FP8Ctx Length:256kTool Calling:SupportedPublished:Jun 11, 2026License:modified-mitArchitecture:Transformer0.8K Open Weights Warm

Kimi K2.7 Code is a 1 trillion parameter Mixture-of-Experts (MoE) model developed by Moonshot AI, with 32 billion activated parameters and a 256K token context length. Built upon Kimi K2.6, this coding-focused agentic model is designed for real-world long-horizon coding tasks and complex software engineering workflows. It features a MoonViT vision encoder with 400M parameters, enabling multimodal capabilities, and improves token efficiency by reducing thinking-token usage by approximately 30% compared to its predecessor. The model excels in coding and agentic benchmarks, offering native INT4 quantization for efficient deployment.

Loading preview...

Kimi K2.7 Code: An Agentic Coding MoE Model

Kimi K2.7 Code, developed by Moonshot AI, is a powerful 1 trillion parameter Mixture-of-Experts (MoE) model with 32 billion activated parameters and a 256K token context length. It is specifically designed as a coding-focused agentic model, building upon the Kimi K2.6 architecture to enhance performance in real-world long-horizon coding tasks and complex software engineering workflows.

Key Capabilities & Features

  • Agentic Coding: Strengthens end-to-end task completion for software engineering, improving token efficiency and reducing thinking-token usage by approximately 30% compared to Kimi K2.6.
  • Multimodal Input: Supports both image and video input through its integrated MoonViT vision encoder (400M parameters), allowing for diverse application scenarios.
  • Mixture-of-Experts Architecture: Features 384 experts with 8 selected per token, contributing to its 1T total parameters while maintaining efficient activation.
  • Native INT4 Quantization: Adopts native INT4 quantization for optimized deployment and inference.
  • Preserve Thinking Mode: Forces preserve_thinking mode by default, retaining full reasoning content across multi-turn interactions to enhance performance in coding agent scenarios.

Performance Highlights

Kimi K2.7 Code demonstrates significant improvements across various benchmarks:

  • Coding: Achieves 62.0 on Kimi Code Bench v2, 53.6 on Program Bench, and 35.1 on MLS Bench Lite, outperforming Kimi K2.6.
  • Agentic Tasks: Scores 46.9 on Kimi Claw 24/7 Bench, 76.0 on MCP Atlas, and 81.1 on MCP Mark Verified, showing strong agentic capabilities.

Good For

  • Complex Software Engineering: Ideal for developers needing assistance with intricate coding tasks and end-to-end software development workflows.
  • Agentic Applications: Suitable for building AI agents that require persistent reasoning and multi-turn interactions.
  • Multimodal Coding: Useful for scenarios where code generation or analysis benefits from visual context (e.g., interpreting diagrams or video instructions).
  • Efficient Deployment: Benefits from native INT4 quantization, making it a strong candidate for resource-conscious deployments using vLLM, SGLang, or KTransformers.

Finetunes
2 models