Skip to content

MoonshotAI/Kimi-K3

  • URL: https://github.com/MoonshotAI/Kimi-K3
  • Stars: 7921
  • Language: Unknown
  • Category: AI & Agents
  • Topics: None

What it does

Kimi K3 is a multimodal AI model designed for tasks involving long-horizon coding, knowledge work, and reasoning. It features a 2.8 trillion parameter architecture and supports a context window of up to one million tokens, allowing it to process and understand text, images, and video simultaneously.

Why it's interesting

Kimi K3 claims to be the world's first open 3T-class model, emphasizing its open-weight nature and advanced architecture, which includes Kimi Delta Attention and a mixture-of-experts approach. Its ability to handle complex tasks with minimal human oversight sets it apart from other AI models.

How it works

The architecture is based on a mixture-of-experts (MoE) framework, activating 16 out of 896 experts for improved efficiency. It includes 93 layers, with a combination of KDA and gated MLA attention mechanisms. The model is designed to perform various tasks, from software engineering to interactive visualizations, leveraging its multimodal capabilities. Specific implementation details about the training process or datasets used are not provided in the README, leaving some uncertainty about its training and evaluation methodologies.

Get started in 5 minutes

To get started, visit the Kimi K3 homepage for documentation and resources. The README does not provide explicit installation instructions or a quick start guide, so further exploration of the repository or linked resources may be necessary.

Watch out for

The repository is relatively new, created in July 2026, which may imply that it is still in development or lacks extensive community support. Additionally, while the model weights are released under the Kimi K3 License, specific licensing terms are not detailed in the README, which could be a concern for potential users.