Skip to content

Kimi models

Moonshot AI's giant open MoE models. Kimi K3 is the first open 3T-class model, with native vision and a 1M-token context.

  • Open weights
  • Models 4
  • Updated Nov 2026
Hugging Face Compare
On this page

At a glance

Type
Open-weight models
Developer
Moonshot AI (China)
First release
2025
Models tracked
4 (3 current)
Abilities
Tool callingReasoning modeImage input
Licence
Kimi K2.x models use a modified MIT licence (attribution for very large products). Kimi K3 and K3.1 Pro use their own Kimi licences; read them before commercial use.
Website
kimi.com
Hugging Face
huggingface.co/moonshotai

Newest model

Kimi K3.1 Pro

Released Nov 2026

Kimi models and versions

ModelReleasedSizesContextAbilitiesLicenceStatus
Kimi K3.1 Pro Most capable Kimi: coding, research and agentsNov 2026 2.8T 2.8T total, 104B active (MoE) Too large for most servers · Ollama cloud—ToolsReasoningVisionKimi K3.1 Pro License Current
Kimi K3 First open 3T-class model, native visionJun 2026 2.8T Too large for most servers · Ollama cloud—ToolsReasoningVisionKimi K3 License Current
Kimi K2.7 Code Coding agent model, ~30% fewer thinking tokens than K2.6Jun 2026 1T Too large for most servers · Ollama cloud—ToolsReasoningVisionModified MIT Current
Kimi K2.6 Multimodal agentic modelApr 2026 1T Too large for most servers · Ollama cloud—ToolsReasoningVisionModified MIT Previous generation

Sizes, context windows and abilities come from the Ollama library and Hugging Face model cards, checked 14 Nov 2026. Release months are the first public upload. Spotted a mistake? Tell us.

What Kimi is good for

Long-horizon coding agents

Kimi K2.7 Code and K3.1 Pro are tuned for multi-hour software tasks.

Hosted open model

Use them through Ollama cloud or an inference provider; self-hosting needs a GPU cluster.

Strengths and limits

Strengths

  • Among the strongest open-weight models for coding and agents.
  • 1M-token context on K3 and K3.1 Pro.

Watch out for

  • Far too large for one server: trillions of parameters.
  • Custom licences on K3 and K3.1 Pro.

About Kimi

Kimi is the model family of Moonshot AI, a Beijing start-up. Kimi K2 (July 2025) was a 1-trillion-parameter mixture-of-experts model released with open weights and became a favourite for coding agents. Later K2 versions added vision (K2.6, April 2026) and a coding-focused variant, Kimi K2.7 Code (June 2026).

Kimi K3 (June 2026) is a new architecture with about 2.8T parameters, Kimi Delta Attention and a 1M-token context; Moonshot calls it the first open 3T-class model. Kimi K3.1 Pro (November 2026) keeps that design, activates 16 of 896 experts (104B active parameters) and is post-trained for coding, research and agent work. These models are open weights, but you will use them hosted unless you run a GPU cluster.

Guides for running Kimi

Tool LLM VRAM Calculator: How Much GPU Memory Do You Need? Updated Nov 2026 AI & LLMs How to Install Ollama on Ubuntu 26.04 and Run LLMs Locally Updated Nov 2026 AI & LLMs How to Install llama.cpp on Ubuntu 26.04 (CPU and NVIDIA GPU) Updated Nov 2026 AI & LLMs How to Install vLLM on Ubuntu 26.04 (OpenAI-Compatible LLM Server) Updated Nov 2026 AI & LLMs How to Install Open WebUI on Ubuntu 26.04 (Self-Hosted ChatGPT Alternative) Updated Nov 2026

Frequently asked questions

What is the latest Kimi model?

The newest Kimi model is Kimi K3.1 Pro, released in Nov 2026.

Can I use Kimi commercially?

Kimi K2.x models use a modified MIT licence (attribution for very large products). Kimi K3 and K3.1 Pro use their own Kimi licences; read them before commercial use. Always read the licence of the exact model you deploy; it is linked in the models table.

Compare Kimi

← All AI models

Back to top