Moonshot AI has introduced Kimi K3, a 2.8-trillion-parameter model designed for long-horizon coding, knowledge work, and complex reasoning. The model features native vision and a 1-million-token context window, enabling it to process images, videos, and documents simultaneously.
Kimi K3 is built on architectural innovations including Kimi Delta Attention and Attention Residuals, which aim to deliver smarter and more efficient scaling. According to Moonshot AI, it is the first open model to reach the 3-trillion-parameter scale, though the exact parameter count is 2.8 trillion.
The model demonstrates capabilities in creative tasks, such as generating playable 3D games and editing videos. In one example, K3 edited its own teaser from 56 source clips, a task that could take an experienced editor one to two working days. It also supports knowledge work, creating professional presentations, interactive research reports, and live dashboards.
For technical work, Kimi K3 can handle chip design, scientific computing, kernel optimization, and compiler development. Moonshot AI reports that K3 designed, optimized, and verified a chip for a nano model autonomously over a 48-hour run using open-source EDA tools.
Kimi K3 is accessible across multiple platforms, including web, mobile, desktop, terminal, and API, and is available through Kimi Web, Kimi Work, Kimi App, Kimi Code, and Kimi API.