NVIDIA has released Cosmos 3, an open physical AI foundation omni-model family designed to combine vision reasoning, world generation, and action prediction. Built on a mixture-of-transformers architecture, the model family enables developers to simulate future states, generate synthetic data, and construct physical AI systems without maintaining separate models for each task.
The foundation models are available under the Linux Foundation’s OpenMDW 1.1 license, allowing organizations to post-train models on their own hardware and proprietary data. NVIDIA stated that physical AI models must understand physical behaviors and predict consequences across robots, autonomous vehicles, and vision systems.
Model architecture and sizes
The Cosmos 3 lineup includes three distinct model sizes built for different computing environments. Cosmos 3 Super contains 64 billion parameters for high-fidelity world modeling, while Cosmos 3 Nano operates with 16 billion parameters to handle reasoning and post-training tasks efficiently.
The smallest variant, Cosmos 3 Edge, uses 4 billion parameters for on-device vision reasoning and robot policy deployment. Designed to run on edge hardware, Cosmos 3 Edge deploys across NVIDIA RTX GPUs, DGX systems, and Jetson platforms, including Jetson Thor.
Benchmark rankings and integrations
In benchmark evaluations, Cosmos 3 achieved the top ranking on Artificial Analysis for open weights text-to-image and image-to-video generation. It also placed first on PAI-Bench for world generation, the Physics-IQ image-to-video category, and RoboLab for robot policy. Cosmos 3 Super recorded the highest score among open models on VANTAGE-Bench for vision understanding.
NVIDIA integrated the new model family into its physical AI software stack, alongside Isaac GR00T for robotics, Alpamayo for autonomous vehicles, and Metropolis for vision AI. Omniverse libraries, provided within the NVIDIA Agent Toolkit, combine with OpenUSD to generate simulation-ready 3D worlds and synthetic data.
Industry adoption and expansion
Several technology companies are building applications using Cosmos 3. Adopters include Doosan Robotics, LG Electronics, Samsung Electronics, and Skild AI in robotics, as well as Li Auto, Xiaomi, and Afari in autonomous vehicle development. Vision AI agents are being developed by Centific, Fogsphere, Linker Vision, Milestone Systems, and Yuan.
NVIDIA also expanded the NVIDIA Cosmos Coalition to Japan, where robotics and manufacturing leaders intend to join and create open world models for factories, logistics, agriculture, construction, healthcare, and transportation.
