NVIDIA unveils Cosmos 3 Edge at SIGGRAPH; 4B-parameter world model for on-device physical AI inference

NVIDIA announced Cosmos 3 Edge at SIGGRAPH 2026 (July 20–23) in Los Angeles: a 4-billion-parameter world model built to run in real time on-device without cloud connectivity. Cosmos 3 Edge joins the broader Cosmos family as an open-weight model designed for embodied AI agents—humanoid robots, grippers, self-driving cars—enabling local prediction and simulation without reliance on external inference infrastructure.

Ming-Yu Liu, VP of Cosmos Lab at NVIDIA, demonstrated that Cosmos uses a mixture-of-transformers architecture to give a single backbone the ability to handle diverse physical tasks: world understanding, prediction, simulation, and action policy. The model compresses different embodiments' representations into a common vocabulary, so robots and autonomous systems can share learned representations despite different hardware. Cosmos-Dreams, a separate simulator framework, generates closed-loop scenarios for testing on a single RTX PRO 6000 GPU before fleet deployment.

Beyond Cosmos, NVIDIA highlighted advances in neural rendering (3D-guided synthesis preserving artistic intent, 4K real-time rendering), physics simulation (NVIDIA Earth-2 weather and climate models achieving 1M-times checkpoint compression), and Model Context Protocol integrations bringing AI agents directly into content-creation tools like Unreal Engine and Maya.

For roboticists and embedded-systems teams: Cosmos 3 Edge removes the inference latency and bandwidth constraint of cloud-dependent models, enabling real-time physical AI at the edge. The open-weight release means teams can fine-tune for domain-specific embodiments. Watch for adoption in warehouse automation, industrial inspection, and consumer robotics where latency and privacy matter.