Cosmos3-Edge, Cosmos3-Nano, and Cosmos3-Super models now available on Amazon SageMaker JumpStart
AWS customers can now access NVIDIA’s Cosmos3 family of omnimodal AI models via Amazon SageMaker JumpStart for robotics, autonomous systems, and vision applications.
Amazon SageMaker JumpStart now hosts NVIDIA’s Cosmos3-Edge, Cosmos3-Nano, and Cosmos3-Super models, expanding AWS’s foundation model catalog for physical AI applications. The Cosmos3 family targets robotics, autonomous vehicles, and vision AI systems requiring real-time reasoning and multimodal input processing. Deployment is streamlined through SageMaker JumpStart, allowing users to integrate these models into workflows with minimal setup via the SageMaker console or Python SDK.
Cosmos3-Edge, a 4-billion-parameter model, is optimized for embedded systems like NVIDIA Jetson Thor, enabling real-time visual reasoning and robot control at 15 Hz with 32 actions per inference. It supports video inputs at 12–30 FPS in 256p and 480p resolutions, making it suitable for edge devices where latency and power constraints are critical.
Cosmos3-Nano, a 16-billion-parameter model, specializes in physics-aware world generation and physical reasoning, processing text, images, video, audio, and action trajectories. It supports chain-of-thought reasoning across modalities up to 720p resolution, aiding robots and vision AI agents in tasks requiring prior knowledge and common-sense reasoning.
Cosmos3-Super, the largest at 64 billion parameters, delivers high-fidelity world generation and simulation across language, images, video, audio, and actions within a unified architecture. It supports resolutions up to 720p and is designed for large-scale simulation, synthetic data generation, and policy learning workflows in AWS environments.