Cosmos 3 is now supported in SGLang-Diffusion.

Cosmos 3 is NVIDIA’s open world model family for Physical AI, combining vision reasoning, world generation, and action-oriented multimodal modeling across text, images, video, audio, and actions.

Serve NVIDIA Cosmos3 generator models (Cosmos3-Nano, Cosmos3-Super, and specialized Super checkpoints) with native SGLang runtime and OpenAI-compatible APIs: