Kimi K3

Summary: Kimi K3 is an advanced open-weight AI model developed by Moonshot AI, designed for complex reasoning, long-horizon coding, multimodal understanding, and knowledge-intensive tasks. Built with a 2.8-trillion-parameter Mixture-of-Experts (MoE) architecture, it features native vision capabilities and a 1-million-token context window. The model supports software development, research, document analysis, visual reasoning, and autonomous AI workflows. Kimi K3 is accessible through Kimi.com, Kimi Work, Kimi Code, and the Kimi API.

What is Kimi K3?

Kimi K3 is a flagship large language model developed by Moonshot AI, a Chinese artificial intelligence company. It is designed to handle demanding AI tasks that require advanced reasoning, long-term context retention, coding capabilities, and multimodal understanding.

The model uses a Mixture-of-Experts architecture with 2.8 trillion total parameters and approximately 104 billion activated parameters. Its architecture combines Kimi Delta Attention (KDA), Attention Residuals (AttnRes), and a Stable LatentMoE framework to improve computational efficiency and information processing.

One of Kimi K3's major capabilities is its 1-million-token context window, allowing it to process extensive documents, large code repositories, research materials, and other information within a single context. Its native vision capabilities also allow it to understand visual inputs and work with image-based tasks.

For software developers, Kimi K3 supports complex coding workflows, repository analysis, GPU kernel optimization, compiler development, and interactive application creation. It can also use visual feedback, such as screenshots, to refine software interfaces and interactive experiences.

Beyond coding, Kimi K3 supports knowledge-intensive workflows such as deep research, scientific analysis, document understanding, and the creation of interactive visualizations and dashboards. Its open-weight availability also gives developers and organizations opportunities to deploy, integrate, and experiment with the model.

Verifiable Stats

1. Developer: Moonshot AI.

2. Model Release: July 2026; the model was introduced publicly in July, with subsequent availability across additional platforms.

3. Model Architecture: Mixture-of-Experts (MoE), using Kimi Delta Attention and Attention Residuals.

4. Total Parameters: 2.8 trillion.

5. Active Parameters: Approximately 104 billion per token.

6. Context Window: 1,048,576 tokens (1 million tokens).

7. Supported Modalities: Text and image; the model's technical materials also describe video understanding capabilities.

8. Core Capabilities: Advanced reasoning, long-horizon coding, visual understanding, deep research, and agentic knowledge work.

9. Availability: Kimi.com, Kimi Work, Kimi Code, and Kimi API.

10. Open Weights: Released under the Kimi K3 License, enabling research, deployment, and further development.