NVIDIA's Cosmos 3 World Models Land on Amazon SageMaker JumpStart

NVIDIA's Cosmos 3 family, three purpose-built world models for physical AI, is now deployable directly from Amazon SageMaker JumpStart. The release covers Cosmos3-Edge, Cosmos3-Nano, and Cosmos3-Super, each aimed at a different point on the spectrum between on-device robot control and high-fidelity simulation.

All three are described as open, frontier omnimodal models designed to help teams build robots, autonomous vehicles, and vision systems that can perceive, reason, plan, and act in physical environments. The differentiator across the lineup is scale and where the compute runs.

Three models, three jobs

Cosmos3-Edge is the smallest at 4 billion parameters, paired with a 2 billion-parameter Nemotron-based reasoner. It's built specifically for real-time robot control on edge hardware, running at 640x360 resolution and generating 32 actions per inference at 15 Hz on NVIDIA's Jetson Thor.

It also handles video at lower resolutions, 256p and 480p, at 12 to 30 frames per second, prioritizing speed over visual detail since it's meant to sit directly in a robot's control loop.

Cosmos3-Nano sits in the middle at 16 billion parameters. It's built for physics-aware world generation and reasoning, processing text, image, video, audio, and action trajectories together, with chain-of-thought reasoning across text, image, and video up to 720p.

Cosmos3-Super is the flagship at 64 billion parameters and the highest-fidelity model in the family. It uses a unified Mixture-of-Transformers architecture to jointly process and generate language, images, video, audio, and action sequences, supporting resolutions up to 720p across multiple aspect ratios.

Why the split matters

The three-tier structure maps cleanly onto how physical AI systems actually get built and deployed. Cosmos3-Super is positioned for training-time simulation and world generation where fidelity matters most, Cosmos3-Nano for reasoning-heavy planning tasks, and Cosmos3-Edge for the constrained, latency-sensitive inference that has to run on the robot itself.

Making all three available through SageMaker JumpStart lowers the integration barrier for AWS customers. Deployment requires only a few clicks through the SageMaker console, or programmatic access via the SageMaker Python SDK, rather than a custom build-out of inference infrastructure.

For robotics and autonomous vehicle teams already working inside AWS, that turns what used to be a multi-week model-serving project into a managed endpoint decision. It also gives NVIDIA another distribution channel for Cosmos models beyond its own NGC and API catalogs, reaching customers who standardize their MLOps stack on SageMaker rather than NVIDIA's own tooling.