Nvidia released Cosmos 3, a family of open-weights world foundation models built for physical AI, spanning three sizes: Cosmos 3 Super at 64B parameters, Cosmos 3 Nano at 16B, and Cosmos 3 Edge at 4B. The family ships under the Linux Foundation's OpenMDW 1.1 license, which permits customers to post-train the models on their own data and hardware. Nvidia says Cosmos 3 currently holds the No. 1 open-weights ranking on five separate benchmarks covering image generation, video generation, world generation, robot policy, and vision understanding.
The benchmark sweep is unusually broad. Cosmos 3 leads Artificial Analysis for open-weights text-to-image and image-to-video, PAI-Bench for world generation, and the image-to-video category of Physics-IQ. On RoboLab, which measures robot policy, Cosmos 3 ranks first. Cosmos 3 Super is separately the highest-ranked open model on VANTAGE-Bench for vision understanding. That is five No. 1 spots across five distinct evaluation regimes, achieved with a single mixture-of-transformers architecture rather than five specialized models stitched together.
The pitch is that physical AI systems, robots, autonomous vehicles, and vision agents need a model that predicts what happens next, not one that only describes what it sees. World models learn how environments behave and generate physically grounded scenarios that teams can use to train and validate their systems before real-world deployment. Nvidia is positioning Cosmos 3 as the backbone: developers can use it as a vision-language model, a world simulator that generates synthetic data, or as the foundation for world-action models that drive robot behavior.
“Physical AI has to understand and predict consequences, not just appearances.”— Ming-Yu Liu, Nvidia, author of the Cosmos 3 announcement
Key facts
- 01Nvidia released Cosmos 3 in three sizes: Cosmos 3 Super at 64B parameters, Cosmos 3 Nano at 16B, and Cosmos 3 Edge at 4B.
- 02Cosmos 3 ranks No. 1 on Artificial Analysis for open-weights text-to-image and image-to-video generation, plus PAI-Bench for world generation.
- 03Cosmos 3 Super leads VANTAGE-Bench for vision understanding among open models, and Cosmos 3 tops RoboLab for robot policy.
- 04Models ship under the Linux Foundation's OpenMDW 1.1 license, permitting post-training on customer data and hardware.
- 05Doosan Robotics, LG, Samsung, Li Auto, Xiaomi, and Skild AI are among the named adopters across robotics and autonomous vehicles.
The three-tier structure maps to a deployment gradient. Cosmos 3 Super at 64B is aimed at high-fidelity data generation and simulation in data centers. Cosmos 3 Nano at 16B is sized for efficient post-training runs when teams want to specialize the model for a particular robot or sensor stack. Cosmos 3 Edge at 4B is small enough to run on Nvidia Jetson platforms including Jetson Thor, meaning the same architecture spans DGX systems, RTX workstations, and on-robot inference.
Nvidia signed the July 2026 open letter titled "Open Weights and American AI Leadership" alongside more than 200 other companies and organizations. The argument in that letter, and echoed in the Cosmos 3 release, is that closed frontier models cannot cover every deployment scenario, particularly in physical AI, where every robot, every factory floor, and every vehicle fleet has its own sensors, tasks, and operating conditions. Access to weights and a license that permits modification is not a marketing feature — it is a technical prerequisite for specialization.
“Specialization is where openness becomes a practical technical requirement.”— Ming-Yu Liu, Nvidia, author of the Cosmos 3 announcement
The company is packaging Cosmos 3 with its broader physical-AI stack: Omniverse libraries for building simulation-ready 3D environments, OpenUSD for composing and exchanging 3D scene data, Isaac GR00T for robotics, Alpamayo for autonomous vehicles, and Metropolis for vision AI. The bet is that developers do not just need a world model — they need the surrounding pipeline for generating synthetic data, running simulations, and testing policies before shipping. Omniverse libraries are now part of the Nvidia Agent Toolkit.
Named adopters span three industries. In robotics: Doosan Robotics, LG Electronics, Samsung Electronics, and Skild AI. In autonomous vehicles: Li Auto, Xiaomi, and Afari. In vision AI for industrial and smart-spaces applications: Centific, Fogsphere, Linker Vision, Milestone Systems, and Yuan. Nvidia has also expanded its Cosmos Coalition to Japan, where robotics and manufacturing companies intend to develop open world models targeting factories, logistics, agriculture, construction, healthcare, and transportation.
The open question is whether an open-weights strategy in physical AI can hold against closed alternatives from firms building vertically integrated robot and vehicle stacks. Nvidia's benchmark leadership is against other open models, not against every proprietary system. Teams that adopt Cosmos 3 still have to invest in the post-training data and evaluation work needed to make a general model perform on their specific robot or vehicle — Nvidia's claim is only that starting from open weights makes that investment tractable.
For Nvidia the strategic logic is straightforward. Every developer that adopts Cosmos 3, Omniverse, and Isaac to build a physical-AI system is a developer whose training, simulation, and inference workloads run on Nvidia hardware end-to-end. Open weights in this frame are not a concession to the open-source community — they are the mechanism by which Nvidia's stack becomes the default substrate for the next wave of robots, cars, and vision agents. The five benchmark wins are the marketing; the OpenMDW license is the moat.
Working on something we should cover, or seeing a story we missed? Send leads, documents, or feedback to hello@aichatdaily.com. For sensitive tips, see our secure tips page for Signal and PGP options.
Spotted an error? Email hello@aichatdaily.com with the URL and the issue, or read our full corrections policy.
