Kimi K2.6
Open-source multimodal agentic model by Moonshot AI
Moonshot AI released Kimi K2.6, an open-source mixture-of-experts model designed for autonomous coding and agent swarms. It processes text, images, and video to execute complex, long-horizon tasks with proactive orchestration capabilities.
- Uses 32 billion activated parameters from 384 experts
- Supports video input as an experimental feature
- Coordinates up to 300 sub-agents for tasks
- Achieves 55.5% accuracy on HLE text subset with tools
- Offers native INT4 quantization support
What it is
Kimi K2.6 is a native multimodal agentic model developed by Moonshot AI. It features a Mixture-of-Experts architecture with 32 billion activated parameters and 61 layers. The model supports text, image, and video inputs, aiming to improve long-horizon coding and autonomous task execution. It can coordinate up to 300 sub-agents for complex workflows, handling tasks from document processing to full-stack web development without constant human oversight.
How it came about
Kimi K2.6 builds upon the architecture of its predecessor, Kimi K2.5. The model shares the same structural design as the previous version, allowing for direct reuse of deployment methods. It introduces enhancements in proactive orchestration and coding-driven design, extending capabilities into visual interface generation and persistent background agent management.
Inside the model
The model uses a Mixture-of-Experts design with 384 total experts, selecting eight per token alongside one shared expert. It employs Multi-Head Latent Attention and SwiGLU activation functions. A MoonViT vision encoder with 400 million parameters handles visual inputs. The system supports a context window of 262,144 tokens and uses native INT4 quantization. For inference, it offers Thinking and Instant modes, with specific temperature and top-p settings recommended for each.
What it can do
Kimi K2.6 demonstrates strong performance in end-to-end coding across languages like Rust, Go, and Python. The model can transform prompts into production-ready interfaces and manage cross-platform operations autonomously. It supports search, code interpretation, and web browsing tools for complex reasoning tasks.
How to use it
Weights are available openly via Hugging Face. Moonshot AI provides an API compatible with OpenAI and Anthropic standards on its platform. Developers can verify deployments using the Kimi Vendor Verifier. The model requires transformers version 4.57.1 or higher and is recommended for use with specific inference engines that support its architecture.
Questions people ask about Kimi K2.6
What modalities does Kimi K2.6 support?
The model accepts text, images, and video as inputs. Video chat is currently an experimental feature available only through the official API. It also integrates tools for search, code interpretation, and web browsing.
How many parameters does the model have?
Kimi K2.6 has 32 billion activated parameters per token. It utilizes a Mixture-of-Experts architecture with 384 total experts, selecting eight for each token plus one shared expert.
Is the model open source?
Yes, Kimi K2.6 is an open-source model. Its weights are available on Hugging Face, and it supports native INT4 quantization. An API is also provided for commercial and developer use.
Good reading and sources
There is no English Wikipedia article on Kimi K2.6 yet, so the facts on this page come from the sources listed above: huggingface.co/moonshotai/Kimi-K2.6. Some are the subject's own pages, so its claims about itself are reported as its claims. EveryLLM wrote the text in its own words, with AI assistance, and compares it with the sources to make sure no passage is copied.
It may contain mistakes, and figures change over time. Nothing here is investment or buying advice. If something is wrong, tell us. Last refreshed 2026-11-14.