Summary
NVIDIA announced Cosmos 3 Edge on July 15–16, 2026, a 4B-parameter edge variant of the Cosmos 3 world model family, designed to run real-time vision reasoning and robot policy generation locally on NVIDIA Jetson Thor platforms (T2000/T3000 modules) and RTX GPUs. Unlike cloud-hosted Cosmos 3 Nano (16B) and Super (64B), Cosmos 3 Edge brings the same world-model capabilities onto the robot itself — eliminating cloud dependency for latency-critical deployment.
Key Contributions
- 4B-parameter Cosmos 3 Edge model, adapts to specific robots, vehicles, and sensors in ~1 day of fine-tuning
- Runs on Jetson Thor T2000/T3000, RTX GPUs, and DGX systems for on-device inference
- Built on NVIDIA Nemotron family; supports on-device vision reasoning and generative robot policy output
- Paired with new Metropolis libraries enabling 6× faster Cosmos-based video intelligence pipelines
- Launch aligned with Japan Physical AI Cosmos Coalition (Fujitsu, FANUC, Hitachi, Kawasaki, Sony, Yaskawa, SoftBank, NEC, etc.)
Significance
Cosmos 3 Edge marks a pivotal shift from cloud-centric world model inference to fully on-robot deployment — enabling real-time closed-loop physical AI without network connectivity, a prerequisite for industrial and field robotics at scale.