NVIDIA Drops Cosmos3-Super to Seed Entire Worlds from a Single Prompt

Cosmos3-Super is a new release from NVIDIA that generates video, images, audio, and even robot action plans from mixed inputs like text, photos, and video clips. It's an omnimodal world model built to simulate and reason about physical environments, from factory floors to autonomous driving scenes. The 64-billion-parameter system handles multiple data types at once, turning a single prompt into coordinated visual, sound, and motion outputs.

NVIDIA built the model to speed up Physical AI work in robotics, self-driving vehicles, and smart spaces. Similar to Cosmos3-Super-Text2Image and Cosmos3-Super-Image2Video They released it under a commercially friendly license on Hugging Face, ready for both research and product development. The architecture blends autoregressive and diffusion transformers so it can generate text and continuous media like video without sacrificing quality.

Multimodal generation and physical AI simulation

Small studios and advanced tinkerers can prototype animations or robotics logic entirely on their own hardware. Privacy-minded teams stay in control of sensitive footage by keeping everything local instead of streaming to cloud APIs. Consolidating multiple modalities into one model also reduces the number of separate tools you need to juggle for complex projects.

Hardware demands and known limitations

NVIDIA notes that Cosmos3-Super lacks a true physics simulator, so outputs can include disappearing objects or physically impossible motion. Longer videos and high-resolution clips still show temporal glitches or object drift in challenging scenes. Running the full model asks for eight H200 or H100 GPUs, though memory offloading can squeeze it onto fewer cards with slower speeds.

"Cosmos3 is a collection of Omnimodal world models capable of generating dynamic, high-quality video, image, audio, and action commands from combinations of text, image, video, and action trajectory inputs." — Source: Hugging Face