Introducing Cosmos 3 Edge

Details

NVIDIA has released Cosmos 3 Edge on Hugging Face, a compact 4-billion-parameter world model built for physical AI systems at the edge. Unlike cloud-dependent robot models, it runs directly on NVIDIA Jetson modules (including the newly announced T2000 and T3000), RTX PRO GPUs, DGX systems, and consumer GeForce RTX GPUs — bringing reasoning capability to memory-constrained hardware. The model serves two roles: as a vision language model (VLM) for understanding scenes and as a world action model (WAM) that predicts outcomes and generates robot actions. It operates at 640×360 robot-control resolution, generates 32 actions per inference on Jetson Thor, and hits 15 Hz real-time control — practical frequencies for factory, warehouse, and hospital robotics. Among all 4B-parameter models, it ranks #1 on VANTAGE-Bench for vision analytics and achieves state-of-the-art for robot policy learning. The dual-tower architecture separates understanding (autoregressive transformer over vision and text tokens) from generation (diffusion transformer over vision, audio, and action tokens), while sharing an underlying representation — enabling a single model to handle scene understanding, future simulation, and action generation in one pass. For robotics and industrial automation teams, this is a credible foundation model for on-device physical AI. The open release on Hugging Face lowers the barrier to experimentation, though the source does not explicitly state the license terms.