NVIDIA Launches Cosmos 3 and Nemotron 3 Ultra to Advance Open-Source Physical AI Ecosystem

World models allow machines to interpret their surroundings, predict how scenes may evolve and select actions, making them a key building block for robotics, autonomous vehicles and smart infrastructure. NVIDIA’s Cosmos 3 platform spans text, images, video, ambient sound and action generation. Cosmos 3 Edge targets local deployment, where reducing reliance on cloud or data-center processing can improve latency, privacy and operational resilience for physical AI systems.

NVIDIA released Cosmos 3 Edge on July 20, 2026, as a 4-billion-parameter open model with a 2-billion-parameter reasoning module that can run independently. On Jetson Thor, it supports a 15-Hz real-time control loop, generating 32 actions per inference at 640-by-360 resolution. The model is optimized for Jetson, RTX PRO, GeForce RTX and DGX systems. Its weights are available on Hugging Face, while NVIDIA has published inference and post-training frameworks and recipes on GitHub.

AI agents currently tend to call separate vision, speech and language models. Transferring data between them adds latency and can result in lost context. NVIDIA has integrated video, audio, image, text and document understanding into a single open model aimed at applications including computer control, customer service and monitoring. The launch has implications for enterprise deployment costs and competition within the AI software ecosystem.

NVIDIA released Nemotron 3 Nano Omni on April 28, 2026. The model uses a 30B-A3B mixture-of-experts architecture, with 30 billion total parameters and about 3 billion activated during inference, and supports a 256K context window. NVIDIA says it can deliver up to nine times the throughput of comparable open models. Seven companies, including Foxconn Technology Group and Palantir, have already adopted it.