TL;DR
- Launch Scope: NVIDIA has launched Cosmos 3 as a physical-AI model that generates both world data and robot-action data for robotics and vision systems.
- Deployment Path: OpenMDW-1.1 gives developers one framework for model artifacts, code, documentation, and data, with access through build.nvidia.com, open repositories, and NIM packaging.
- Backdrop: The release extends NVIDIA’s physical-AI strategy, while Marble and Genie remain older, hedged comparisons rather than proof of today’s rollout.
NVIDIA launched Cosmos 3 for physical AI this week at GTC Taipei during Computex as a model built to generate both world data and robot-action data for machines operating in changing real-world scenes. Robots, autonomous vehicles, and large vision systems are the intended targets, giving the release practical stakes beyond a research demo.
Open-source packaging is part of the rollout, too. The Linux Foundation released OpenMDW-1.1 on May 28 as a framework for AI model distributions, and developers can use that framework to keep model artifacts, code, documentation, and data under one legal structure.
With the single model-centric license, developers can train, modify, contribute, redistribute, and deploy weights, architecture, documentation, datasets, benchmarks, and code without splitting them across separate legal bundles.
One on-page NVIDIA caption describes the product as “Cosmos 3 powers perception, prediction and action.” NVIDIA is trying to position physical AI as deployable engineering software rather than another chatbot-style model.
Cosmos 3 Turns Scene Understanding Into Action
Cosmos 3 is a world model, meaning an AI system designed to simulate how an environment changes rather than simply label what a camera sees. Its reasoning block interprets a scene before a generation stage produces grounded outputs such as synthetic video and robot-task data.
Native action generation pushes the system beyond visual prediction. Cosmos 3 can emit numerical robot data such as joint angles, gripper positions, and trajectory points, giving robotics teams material they can feed into planning and control workflows.
Cosmos 3 can also generate physically plausible video sequences for rare or costly situations. That gives teams artificial training material for robot training and world-model simulation.
Developers also get immediate access paths. Teams can try Cosmos 3 on build.nvidia.com, download open models from Hugging Face and GitHub, and use those resources for direct experimentation.
Deployment is part of the rollout rather than an afterthought. NVIDIA also positions Cosmos 3 for packaging through NVIDIA NIM microservices, which gives customers a cleaner path from testing into production services.
Early Deployments and Product Readiness
Named deployments are part of NVIDIA’s case that Cosmos 3 is meant for active engineering use. NVIDIA’s GEAR team is using it to develop video action models for embodied agents across games, simulations, and robotics environments.
Agile Robots is using action-conditioned robot data to build policy-development workflows and generate task trajectories at scale.
Linker Vision is using Cosmos-based reasoning to analyze live camera streams and perform root-cause analysis across large video networks.
Licensing also extends beyond the first model drop. Future Cosmos-family open models will adopt OpenMDW-1.1.
Cosmos Super and Cosmos Nano are available now, while Cosmos Edge is expected later for real-time inference. Cosmos 3 Nano and Cosmos 3 Super are the 8B and 32B variants, with Nano tuned for efficient inference and Super aimed at higher-quality, larger-scale deployment.
NVIDIA also launched the Cosmos Coalition with Agile Robots, Black Forest Labs, Generalist, LTX, Runway, and Skild AI as founding members. That list suggests the company wants the model family positioned as shared infrastructure for a broader physical-AI ecosystem, not only as an in-house research project.
NVIDIA’s DreamDojo robot-training work had already pointed toward the same robotics direction before this launch. A GTC 2025 reasoning-model push before tied Cosmos to NVIDIA’s broader open-model strategy.
