NVIDIA released Cosmos 3, a foundation model that brings together physical reasoning, world generation, and action generation for robotics and autonomous systems. Unlike previous Cosmos versions that required separate models for each capability, this release uses a Mixture-of-Transformers architecture with distinct Reasoner and Generator towers to handle multimodal understanding and physics-aware output creation.

Cosmos 3 Nano packs 16B parameters for workstation deployment on RTX PRO 6000 GPUs, while the 64B Cosmos 3 Super targets datacenter-scale applications on Hopper and Blackwell hardware. Both models are available as open-source checkpoints on Hugging Face alongside training scripts and deployment tools, making physical AI development more accessible to research teams.

The release includes six synthetic datasets covering robotics, autonomous driving, and warehouse operations, plus the NVIDIA Cosmos Human Evaluation framework for objective model assessment. Cosmos 3 leads public benchmarks including VANTAGE-Bench and PAI-Bench, demonstrating strong performance across physical reasoning and video generation tasks.

Teams can now build end-to-end physical AI systems without orchestrating multiple models, potentially accelerating development of manipulation robots, autonomous vehicles, and smart environment monitoring solutions.