NVIDIA Cosmos

NVIDIA's open world-foundation-model platform, using video generation to produce training data and rehearsal environments for robots and self-driving cars.

Cosmos is the world-foundation-model platform NVIDIA released at CES on January 6, 2025. A world foundation model (WFM) is a general-purpose model first pretrained on massive amounts of video to learn “what happens next in the world,” then post-trained for a specific robotics or self-driving scenario. The platform includes a video-curation pipeline (cutting about 20 million hours of raw video down to roughly 100 million clips), the Cosmos Tokenizer video tokenizer, both diffusion-based and autoregressive pretrained models, and post-training examples, all released with open weights. Several product lines followed: Cosmos Predict, which predicts future frames; Cosmos Transfer, which generates photorealistic video conditioned on structured inputs like segmentation or depth maps; and Cosmos Reason, a multimodal large model for physical commonsense and embodied reasoning. Cosmos 3, from June 2026, folds language, image, video, audio, and action into a single mixture Transformer model, offered in Edge, Nano (16B), and Super (64B) sizes, released under the Linux Foundation's OpenMDW-1.1 license.

ExampleOne use of Cosmos Transfer is sim-to-real: taking segmentation maps, depth maps, and edge maps rendered by a simulator as conditioning, it generates video that looks like real footage, used to expand training data for robots and self-driving cars.

Also called

Cosmos World Foundation Models, Cosmos WFM

Related

World Foundation Model · NVIDIA Cosmos Predict · NVIDIA Cosmos Transfer · NVIDIA Cosmos Reason · Cosmos 3 · NVIDIA

Sources

Cosmos World Foundation Model Platform for Physical AI (arXiv 2501.03575)
NVIDIA Launches Cosmos World Foundation Model Platform to Accelerate Physical AI Development (NVIDIA Newsroom, 2025-01-06)
Cosmos 3: Omnimodal World Models for Physical AI (arXiv 2606.02800)

As of

2026-06