Cosmos Predict-1

Cosmos, announced by NVIDIA at CES in January 2025, is a platform of generative world foundation models (WFMs) for physical AI — models that predict and generate physics-aware video of virtual environments so robots and autonomous vehicles can be trained on synthetic data instead of expensive real-world collection. The Predict-1 family (diffusion and autoregressive variants in Nano, Super, and Ultra sizes) handles text-to-world and video-to-world generation.

The models are open-weight under NVIDIA's permissive open model license (commercial use allowed), trained on a reported 9,000 trillion tokens from 20 million hours of real-world video, and downloadable from NGC, Hugging Face, and the NVIDIA API catalog. Siblings include Cosmos-Transfer1 (controllable generation from depth/segmentation inputs) and Cosmos-Reason (chain-of-thought video reasoning). Early adopters included 1X, Agility, Waabi, Uber, and XPENG.

Honest caveats: these are world simulators, not chat models — they generate video, not actions — and serious fine-tuning wants NVIDIA GPUs. The '9,000 trillion tokens' figure is NVIDIA's. For robotics and AV teams, Cosmos is the closest thing to an open foundation-model stack for physical AI.