2.8T Parameters
First open model to reach the 3T-parameter class
Kimi K3 is an open 2.8T-parameter model built on our Kimi Delta Attention and Attention Residuals, with native vision and a 1M-token context window, designed for long-horizon coding, knowledge work, and reasoning. It turns a prompt into playable games, polished presentations, and interactive research reports, even taking on technical work once reserved for experts.
More scale, longer context, native vision, and a smarter architecture, all built to push Kimi further.
First open model to reach the 3T-parameter class
Built for long conversations and complex projects
Understands images, videos, and documents together
Built on Kimi Delta Attention and Attention Residuals for smarter, more efficient scaling
From simple prompts to finished creative work.
Powered by K3's advanced spatial reasoning, coding, and visual understanding, Kimi can turn descriptions, images, and videos into fully playable interactive experiences. It researches references, builds 3D assets, writes the game logic, then looks at live screenshots of its own build and refines the result until it plays right.
An explorable open world of forests, a log-cabin village, snowy mountains, and shifting weather. External assets used: animated cowboy and horse models and terrain data.
With its native multimodal architecture, Kimi K3 understands text, images, and video together. It can select the right clips, match motion between cuts, synchronize edits to the beat, and process audio. K3 edited its own teaser from 56 source clips, a high-density edit that can take an experienced editor one to two working days, or a beginner three to five.
K3 turns complex information into polished slides, interactive reports, and live dashboards you can explore and use.
Kimi K3 builds professional presentations through in-depth research and complex analysis, with custom charts, scientific diagrams, and infographics, all fully editable and tailored to the content. From industry research to scientific analysis, see how K3 makes complex ideas clear and presentation-ready.
With K3, large-scale research becomes an interactive report you can explore and drill into, with evidence-backed insights brought to life through bespoke charts, animated diagrams, and visual narratives. See it in action in a 42-year deep dive into the AI ASIC industry.
Move beyond one-off answers with a live dashboard built from widgets that connect to your data and keep updating. Bring the ones that matter together around your project or goal, so your work stays organized, visible, and easy to return to. Available in Kimi Work.
Kimi K3 stays with complex technical work for longer, moving between code, tools, testing, and iteration with less human oversight. From chip design to scientific computing, kernel optimization, and compiler development, see how far K3 can take a long-running technical task.
K3 designed, optimized, and verified a chip for a nano model built on its own architecture, completing the full workflow autonomously with open-source EDA tools.A chip built by a model, for a model · 48-hour autonomous run
Kimi K3 delivers frontier-level performance across demanding coding, tool-use, reasoning, and knowledge-work benchmarks. Explore what it can do, or dive into the full benchmark results and evaluation details in the Tech Blog.
Kimi K3 works wherever you do, from the web and mobile app to your desktop, terminal, and API. Pick the way that fits how you work.
Kimi K3 can be used in Kimi Code from the terminal or supported IDEs. In Kimi Code CLI, use the /model command to select Kimi K3. In Kimi Code for VS Code and Cursor, select Kimi K3 from the model dropdown in the input bar.
In Kimi Work, you can create interactive Widgets directly in a conversation and connect them to local data or external plugins for ongoing updates. You can then add the Widgets you want to keep to Dashboard, where they remain available in a persistent view organized around a topic, project, or goal. This makes it easier to keep frequently used information and tools in one place and return to them across different tasks.
Kimi K3 is an open-weight, native multimodal model developed by Moonshot AI. It has 2.8 trillion total parameters and is built on Kimi Delta Attention (KDA) and Attention Residuals (AttnRes), with native vision capabilities and a context window of up to 1 million tokens. It is designed for long-horizon coding, knowledge work, and reasoning.
Yes. Kimi K3 is open source. Its full model weights are available on Hugging Face, and both the model weights and code are released under the Kimi K3 License. The license allows users to run, deploy, fine-tune, and modify the model, subject to its terms and conditions.
Kimi K3 is available through Kimi Web, the Kimi mobile app, Kimi Work, Kimi Code, and the Kimi API. The Kimi app is available on iOS, Android, and HarmonyOS. Organizations can use Kimi through Kimi Business, and developers can also download the model weights from Hugging Face for their own deployments.
Kimi K3 supports a context window of up to 1 million tokens, allowing it to keep more information in context over long conversations and complex tasks.
Yes. Kimi K3 is a native multimodal model that can understand text, images, and video within the same model. This allows it to work with visual information such as screenshots, documents, and video alongside text.