NVIDIA has released Cosmos 3, an open family of world models aimed at advancing physical AI. The models show leading benchmark results and are adopted in robotics, autonomous vehicles, and vision AI. The article emphasizes the importance of open models in physical AI and the role of NVIDIA Omniverse libraries in building simulation-ready worlds.
SynthePulse Insight · AI deep reading
Version 1 · 1 source
NVIDIA releases Cosmos 3, using open weights and a unified architecture to drive the adoption of physical AI, but balancing openness and performance remains a key challenge.
In July 2026, NVIDIA signed the 'Open Weights and American AI Leadership' open letter, arguing that AI leadership depends on whether the open ecosystem covers every industry, not just a single frontier model. This stance directly drove the release of Cosmos 3, emphasizing the necessity of open models in physical AI.
Physical AI needs to understand and predict the consequences of physical environments, not just their appearance. World models learn the behavior of physical environments, predict future states, and generate physics-based data, providing a customizable foundation for robotics, autonomous driving, and visual AI.
Cosmos 3 is a frontier open physical AI foundation omni-model, based on a hybrid Transformer architecture that combines visual reasoning, world generation, and action prediction, so developers don't need to maintain separate models for each capability.
The family includes Cosmos 3 Super (64B) for high-fidelity world modeling, Cosmos 3 Nano (16B) for efficient inference and post-training, and Cosmos 3 Edge (4B) for visual reasoning and robot policy deployment on edge devices. The Edge version runs on RTX GPUs, DGX systems, and Jetson (including Jetson Thor).
In benchmarks, Cosmos 3 ranks first in Artificial Analysis for open-weight text-to-image and image-to-video generation, in PAI-Bench for world generation, in Physics-IQ for image-to-video, in RoboLab for robot policy, and Cosmos 3 Super ranks highest in VANTAGE-Bench for visual understanding.
The Cosmos world foundation models use the Linux Foundation's OpenMDW 1.1 license, allowing teams to post-train on their own data and hardware. This openness is seen as a practical technical requirement for specialization, because general models cannot cover every team's specific robots, sensors, or operating environments.
Beyond model specialization, teams also need environments to generate data, run simulations, and test behaviors. Omniverse libraries and OpenUSD provide an open framework for composing, reusing, and exchanging complex 3D data, reducing redundant work when assets, sensor configurations, or environmental conditions change.
Cosmos has been adopted by several companies: in robotics, including Doosan Robotics, LG Electronics, Samsung Electronics, and Skild AI; in autonomous driving, including Li Auto, Xiaomi, and Afari; and in visual AI, including Centific, Fogsphere, Linker Vision, Milestone Systems, and Yuan.
The NVIDIA Cosmos Alliance brings together world model builders, AI developers, and physical AI leaders, recently expanding to Japan, attracting robotics and manufacturing leaders to develop open world models for factories, logistics, agriculture, construction, healthcare, and transportation.
This article's information primarily comes from NVIDIA's official blog, which is vendor-published content. Benchmark results and adoption cases have not been independently verified and should be treated with caution.
Cosmos 3 attempts to lower the barrier to physical AI specialization through open weights and a unified architecture, but its actual effectiveness and ecosystem influence still require more independent verification.
Primary report
Primary source