NVIDIA releases Cosmos 3 Edge, a 4-billion-parameter open world model for robotics and vision AI on edge devices

NVIDIA announced the release of Cosmos 3 Edge, a 4-billion-parameter open world model designed for robotics and vision AI agents operating on edge devices. The model is available on Hugging Face’s Cosmos 3 repository and targets memory-constrained systems such as NVIDIA RTX PRO GPUs, NVIDIA DGX, GeForce RTX GPUs, and NVIDIA Jetson modules, including the newly announced Jetson T2000 and T3000.

Cosmos 3 Edge is engineered to deliver memory-efficient, high-throughput inference while supporting real-time reasoning and action generation. On NVIDIA Jetson Thor, the model processes 640×360 observations and generates 32 actions per inference, achieving real-time control at 15 Hz. Among models with approximately 4 billion parameters, it ranks first on VANTAGE-Bench for vision analytics and sets state-of-the-art performance for robot policy learning, according to the announcement.

The model architecture combines two transformer towers: an autoregressive tower for understanding and reasoning over vision and text tokens, and a diffusion tower for prediction, generation, and neural simulation across vision, audio, and action tokens. The towers share multimodal attention layers but maintain separate normalization and MLP layers, enabling a shared representation for scene understanding, future prediction, and action planning.

Cosmos 3 Edge introduces a common action representation that maps diverse physical embodiments—such as vehicles, cameras, robot arms, and grippers—into compact geometric vectors encoding translation, rotation, and manipulation state. This unified representation allows the model to link visual changes with physical motion and control inputs, enabling developers to generate actions grounded in cause and effect.

The release includes a policy mode where the model predicts an action alongside its expected visual consequence, connecting world modeling directly to robot policy training. NVIDIA also provides Cosmos 3 Edge Policy (DROID), a robot manipulation policy post-trained on the DROID dataset for pick-and-place tasks, along with post-training scripts for fine-tuning on small clusters of H100 or NVIDIA DGX Station before deployment to edge platforms.

NVIDIA is distributing post-trained checkpoints and training recipes to help developers adapt Cosmos 3 Edge to domain-specific workloads. The company also released the Cosmos 3 Super 4-Step Distillation checkpoint, which reduces diffusion steps from 35–50 to 4, enabling up to 25× faster inference while preserving image and video quality. These artifacts are provided as reference implementations on Hugging Face.