NVIDIA releases Cosmos 3 Edge, a 4B open world model for robots on Hugging Face

NVIDIA released Cosmos 3 Edge, the latest model in its Cosmos world-model family, publishing it openly on Hugging Face. The 4-billion-parameter model is designed for robots and vision AI agents, optimized for memory-efficient, high-throughput inference on NVIDIA edge hardware — including Jetson Thor, RTX PRO, DGX, GeForce RTX, and Jetson-class GPUs — to enable real-time environmental understanding and action generation on resource-constrained devices.

Cosmos world models predict physically plausible future frames and dynamics, serving as the perception-and-simulation backbone for physical AI and robotics. Bringing a compact 4B variant to the edge extends those capabilities out of the datacenter and onto devices, where latency and power constraints matter most. NVIDIA researcher Liu Mingyu highlighted the availability on Jetson Thor, and the release was echoed on Hugging Face's own blog.

The launch fits NVIDIA's broader physical-AI push showcased at SIGGRAPH 2026, where the company advanced graphics, real-time simulation, and agentic and physical AI, and detailed NVLink as the scale-up network for AI factories. It also aligns with NVIDIA and Hugging Face's expanded LeRobot open-robotics collaboration.

Strategically, offering open edge models keeps NVIDIA's hardware central to the fast-growing robotics and embodied-AI market — a segment where Mistral (Robostral Navigate) and others are also competing. The open release lowers adoption friction while ensuring the models run best on NVIDIA silicon. Readers should watch benchmark comparisons against other embodied/world models, real-world robot deployments, and how the compact edge variant performs versus larger cloud-hosted Cosmos models.