4B Cosmos 3 Edge Runs Real-Time Robot Action on Jetson Thor

NVIDIA launched Cosmos 3 Edge, a 4-billion-parameter open world model engineered for real-time physical AI and robotics inference. The release completes the Cosmos 3 family, joining the existing 8B Nano and 32B Super variants, but targets local reasoning on memory-constrained edge hardware.

Built on the NVIDIA Nemotron architecture, the model operates as a unified World-Action Model (WAM). It drops standard Vision-Language Model paradigms to combine visual reasoning, environment prediction, and action generation using a mixture-of-transformers backbone. Multimodal attention is shared across autoregressive and diffusion transformer towers.

Performance and Inference Constraints

Cosmos 3 Edge is optimized for the NVIDIA Jetson platform, specifically the new T2000 and T3000 modules. The model processes 640×360 control resolution observations and features a built-in 2B Nemotron-based reasoner to predict physical outcomes before committing to movement.

• Throughput: Generates 32 actions per inference.
• Latency: Achieves real-time control at 15 Hz on Jetson Thor.
• Deployment: Supports Jetson hardware, NVIDIA RTX PRO GPUs, DGX systems, and consumer GeForce RTX GPUs.

In early testing on the VANTAGE-Bench vision analytics and robot policy evaluation suite, Cosmos 3 Edge ranked first among 4B-class models.

Strategic Rollout and Licensing

NVIDIA released the weights and inference framework under the OpenMDW-1.1 License, allowing commercial use with attribution. The model is integrated with the NVIDIA Metropolis stack for agentic vision and NVIDIA Isaac for robotics simulation. Developers can access optimized high-throughput inference through vLLM-Omni.

This release aligns with the NVIDIA Cosmos Coalition, a physical AI initiative announced in Tokyo involving FANUC, Yaskawa Electric, Kawasaki Heavy Industries, Sony, SoftBank, and Fujitsu. While Cosmos 3 Edge can be emulated immediately using JetPack, the target Jetson T2000 and T3000 physical modules will not ship until Q1 2027.

If you build multi-agent systems for hardware, Cosmos 3 Edge provides a cloud-free path to visual reasoning. Developers can use the Cosmos open framework to adapt the base model to custom robot embodiments and sensor configurations in roughly one day.