NVIDIA's open platform for Physical AI: World Foundation Models (Cosmos Predict, Transfer, Reason, Cosmos 3) plus pretraining, post-training and evaluation frameworks.

NVIDIA Cosmos is a platform developed by NVIDIA for Physical AI — robotics, autonomous vehicles and industrial vision systems. Cosmos combines a library of open World Foundation Models (WFM) with frameworks for data processing, training, post-training and evaluation.
Cosmos family models: Cosmos Predict (generating future world states from text, image and video), Cosmos Transfer (controllable sim-to-real translation, for example from NVIDIA Omniverse to photorealistic synthetic data), Cosmos Reason (a vision-language model for reasoning over a scene and intent) and Cosmos 3 (a Mixture-of-Transformers omni-model announced at COMPUTEX 2026 that merges perception, reasoning and generation).
Cosmos Curator — filtering, annotation and deduplication of large sensor datasets. Cosmos Tokenizer1 — a video tokenizer for compression and decoding in latent space. Cosmos Evaluator — review and scoring of generative video outputs. Post-training and evaluation frameworks and inference scripts are released in the NVIDIA/Cosmos repository. The Cosmos Cookbook ships ready recipes for robotics, AV and industrial vision.
Weights and reference code are available on Hugging Face (NVIDIA Cosmos1/2/2.5/3 collections and Cosmos-Predict, Cosmos-Transfer, Cosmos-Reason) and on GitHub (NVIDIA/Cosmos). Models can also be run through NVIDIA NIM and from the build.nvidia.com catalogue. Earlier releases were under the NVIDIA Open Model License; Cosmos 3 is released under the OpenMDW 1.1 license from the Linux Foundation.
Cosmos is optimised for NVIDIA GPUs: recommended server platforms with RTX PRO 6000 Blackwell, GB200 Grace Blackwell and Hopper (H100). Together with NVIDIA Omniverse (a 3D simulation environment) it forms a consistent pipeline: Omniverse generates simulated scenes, Cosmos turns them into photorealistic synthetic data and world models used for post-training of robot and vehicle policies. Customers cited by NVIDIA include 1X Technologies, Agility Robotics, Figure AI, Neura Robotics, Toyota Research Institute, General Motors, Uber, Li Auto, Foretellix, Parallel Domain and others.
Real-time collaboration and synchronization mechanism in NVIDIA Omniverse. The Nucleus server manages OpenUSD data and distributes changes to connected clients in a publish/subscribe model (Live Sync / Live Session), letting multiple applications and users work on the same 3D scene simultaneously.