Kimi K3
An open-weight, 2.8-trillion-parameter multimodal mixture-of-experts model built for long-horizon coding, agentic knowledge work and reasoning, with a one-million-token context window.
Kimi K3 is an open-weight, native multimodal agentic model built on a new Kimi Delta Attention and Attention Residuals architecture, with 2.8 trillion total parameters, 104 billion activated per token across 896 experts, and native vision capabilities alongside a one-million-token context window.
It is designed for long-horizon coding sessions across large repositories and terminal tool orchestration, from GPU kernel optimization to chip design, and for agentic knowledge work such as producing research reports with interactive visualizations and motion design, with a text-and-image multimodal architecture.
Its full model weights are released under the Kimi K3 License for research, deployment and further development.