ai/kimi-k2.6 repository overview

⁠Kimi K2.6

Kimi K2.6 is an open-source, native multimodal agentic model developed by Moonshot AI that advances practical capabilities in long-horizon coding, coding-driven design, proactive autonomous execution, and swarm-based task orchestration. Built on a Mixture-of-Experts (MoE) architecture with 1 trillion total parameters and 32 billion activated parameters per token, K2.6 represents a significant step forward in AI-powered development and automation.

The model excels at complex, end-to-end coding tasks with reliable generalization across multiple programming languages including Rust, Go, and Python, spanning domains from front-end development to DevOps and performance optimization. K2.6 is capable of transforming simple prompts and visual inputs into production-ready interfaces and lightweight full-stack workflows, generating structured layouts, interactive elements, and rich animations with deliberate aesthetic precision. It features native multimodal understanding with support for text, image, and video inputs through its custom MoonViT vision encoder.

A standout capability of Kimi K2.6 is its agent swarm technology, which can scale horizontally to 300 sub-agents executing 4,000+ coordinated steps. This enables dynamic task decomposition into parallel, domain-specialized subtasks, delivering end-to-end outputs from documents to websites to spreadsheets in a single autonomous run. The model also demonstrates strong performance in powering persistent, 24/7 background agents that proactively manage schedules, execute code, and orchestrate cross-platform operations without human oversight.

⁠Characteristics

⁠Using this model with Docker Model Runner

docker model run aistaging/kimi-k2.6

For more information, check out the Docker Model Runner docs⁠.

⁠Benchmarks

⁠Agentic Benchmarks

⁠Coding Benchmarks

⁠Reasoning & Knowledge Benchmarks

⁠Vision Benchmarks

Note: All Kimi K2.6 results are with thinking mode enabled. Asterisks (*) indicate results re-evaluated under the same conditions as K2.6.

⁠Key Features

⁠Long-Horizon Coding

• Demonstrates significant improvements on complex, end-to-end coding tasks
• Reliable generalization across Rust, Go, Python and other languages
• Spans domains including front-end, DevOps, and performance optimization
• Capable of executing 4,000+ tool calls over 12+ hours of continuous execution
• In real-world tests, achieved performance improvements such as 185% throughput gains in financial matching engines

⁠Coding-Driven Design

• Transforms simple prompts and visual inputs into production-ready interfaces
• Generates structured layouts, interactive elements, and rich animations
• Creates lightweight full-stack workflows with aesthetic precision

⁠Agent Swarm Technology

• Scales horizontally to 300 sub-agents executing 4,000 coordinated steps
• Dynamically decomposes tasks into parallel, domain-specialized subtasks
• Delivers end-to-end outputs from documents to websites to spreadsheets in a single autonomous run

⁠Multimodal Understanding

• Native support for text, image, and video inputs
• Powered by custom MoonViT vision encoder (400M parameters)
• 256K token context window for processing long documents and conversations

⁠Thinking Mode

• Features explicit reasoning capabilities with visible thought process
• Supports "preserve thinking" mode for multi-turn interactions
• Can be toggled between thinking and instant modes

⁠Architecture Details

Kimi K2.6 uses an advanced Mixture-of-Experts architecture with:

• 61 layers total (60 MoE layers + 1 dense layer)
• 384 experts with 8 experts activated per token
• Multi-Latent Attention (MLA) mechanism for efficient attention computation
• SwiGLU activation function in feed-forward networks
• Native INT4 quantization for efficient deployment

⁠Links

• Official Blog Post⁠
• Hugging Face Model⁠
• Kimi Chat Interface⁠
• API Platform⁠
• Kimi Code CLI⁠
• Technical Paper⁠

⁠Considerations

• The model requires significant computational resources with 1T total parameters, though only 32B are activated per token
• Native INT4 quantization is available for more efficient deployment
• Recommended temperature is 1.0 for Thinking mode and 0.6 for Instant mode
• Recommended top_p is 0.95
• Video input is an experimental feature and may have limitations
• Context management strategies may be needed for tasks requiring very long context
• The model performs best with the Kimi Code CLI framework for coding agent tasks
• Version requirement for transformers is >=4.57.1, <5.0.0
• Currently recommended inference engines are vLLM, SGLang, and KTransformers
• Background agent tasks can run for extended periods (12+ hours) and require appropriate infrastructure

⁠Generated by

This model card was automatically generated using cagent-action⁠. Want to learn more about Docker Model Runner? Check out the project repository: https://github.com/docker/model-runner⁠.