Kimi K3
Open-weight multimodal agentic model with 2.8T parameters
Moonshot AI released Kimi K3 in 2026 as an open-weight multimodal model. It features 2.8 trillion parameters and a one-million-token context window for complex coding and reasoning tasks.
- Contains 2.8 trillion total parameters
- Activates 16 of 896 experts per token
- Supports a 1-million-token context window
- Uses MXFP4 weights for efficiency
- Processes text, images and video natively
What it is
Kimi K3 is a large language model developed by Moonshot AI. It contains 2.8 trillion parameters and uses a mixture of experts architecture. The model processes text, images, and video natively. It supports a context window of one million tokens. This design allows it to handle long engineering sessions and complex knowledge work. The system operates with minimal human oversight for coding and terminal tasks.
How it came about
Moonshot AI introduced Kimi K3 as its most capable model to date. It follows the earlier Kimi K2 version. The company claims this release represents the first open model in the three-trillion parameter class. The development focused on improving scaling efficiency through a new sparse expert framework.
Inside the model
The architecture relies on Kimi Delta Attention and Attention Residuals. It employs a Stable LatentMoE framework that activates 16 experts from a pool of 896. This approach yields a claimed 2.5-fold improvement in scaling efficiency over its predecessor. The model uses MXFP4 weights and MXFP8 activations via quantization-aware training. This setup aims for broad hardware compatibility during inference.
What it can do
The model performs well in long-horizon coding and repository navigation. It handles tasks such as GPU kernel optimization and chip design. Benchmark results include a score of 67.3 on DeepSWE using the mini-SWE-agent harness. On BrowseComp, it achieved 90.4 when using its full context window without management strategies. It also supports interactive visualizations and video editing workflows.
How to use it
Moonshot AI releases the full model weights under the Kimi K3 License. This permits research and deployment use. An API is available through the platform.kimi.ai website. The interface supports OpenAI and Anthropic compatible calls. Recommended inference engines include vLLM, SGLang, and TokenSpeed for local or hosted execution.
Questions people ask about Kimi K3
How many parameters does Kimi K3 have?
The model has a total of 2.8 trillion parameters. It uses a mixture of experts design where only 16 out of 896 experts are active for any given token.
What types of data can Kimi K3 process?
It is a native multimodal model. It can understand and process text, images, and video within the same architecture without needing separate encoders.
Is Kimi K3 open source?
Yes, Moonshot AI releases the full model weights under the Kimi K3 License. This allows for open research, deployment, and further innovation by the community.
What is the context length of Kimi K3?
The model supports a context window of one million tokens. This allows it to process very large documents or code repositories in a single session.
Good reading and sources
There is no English Wikipedia article on Kimi K3 yet, so the facts on this page come from the sources listed above: huggingface.co/moonshotai/Kimi-K3. 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.