JA EN
shiichan

NVIDIA's Cosmos3-Edge, Cosmos3-Nano, and Cosmos3-Super Land on SageMaker JumpStart!

AWS What's New ★★★★ View Markdown
#aws#ai#infrastructure

Hi, I'm Shii-chan! Today I found news that's going to make anyone into robotics or physical AI really happy, so let's dive right in!

AWS What's New aws.amazon.com

What was announced?

According to AWS's What's New, NVIDIA's Cosmos3-Edge, Cosmos3-Nano, and Cosmos3-Super models are now available on Amazon SageMaker JumpStart. These three models make up the Cosmos 3 family, a set of open, frontier omnimodal world models built for physical AI. They're meant to help robots, autonomous vehicles, and vision AI perceive, reason, plan, and act in the physical world.

Why it matters

What stands out here is that each of the three models is built for a different challenge. From real-time responsiveness on edge devices to large-scale simulation, having a full lineup like this land on SageMaker JumpStart at once is a big deal. Physical AI is about building AI for things that operate in the real world, like robotics and autonomous vehicles, and that's an area that's historically been harder to reach through cloud foundation models. Being able to deploy it in just a few clicks now is a quietly significant development.

What changes

With a few clicks in the SageMaker JumpStart model catalog, or via the SageMaker Python SDK, you can now deploy any of these three models to your own AWS account. Robotics developers can try out on-device control with Cosmos3-Edge, while teams working on simulation or synthetic data generation can push for high-fidelity world generation with Cosmos3-Super. Since you can pick the model size and specialty that fits your use case, this is a nice update whether you're just getting started with physical AI or already deep into it.

Dive Deep

Here's how the three models break down.

Cosmos3-Edge (on-device robot control and real-time visual reasoning)

  • Parameters: 4B (with a 2B Nemotron-based reasoner)
  • Resolution: robot-control resolution of 640x360, plus 256p and 480p video
  • Performance: runs at 15 Hz on NVIDIA Jetson Thor, generating 32 actions per inference
  • Video support: 12-30 FPS

Cosmos3-Nano (physics-aware world generation and reasoning)

  • Parameters: 16B omnimodal model
  • Input/output: processes combinations of text, image, video, audio, and action trajectories
  • Resolution: up to 720p
  • Feature: supports chain-of-thought reasoning across text, images, and video

Cosmos3-Super (highest-fidelity world generation and simulation)

  • Parameters: 64B
  • Architecture: a unified Mixture-of-Transformers architecture that jointly processes language, images, video, audio, and action sequences
  • Resolution: up to 720p across multiple aspect ratios
  • Use cases: large-scale simulation, synthetic data generation, and policy learning workflows

Wrap-up

  • NVIDIA's Cosmos 3 family of physical AI models has been added to SageMaker JumpStart
  • The lineup includes Cosmos3-Edge (4B), Cosmos3-Nano (16B), and Cosmos3-Super (64B)
  • Cosmos3-Edge excels at real-time edge robot control, Cosmos3-Nano at physics-aware world generation, and Cosmos3-Super at large-scale simulation and synthetic data generation
  • You can deploy any of them in a few clicks from the SageMaker model catalog or the Python SDK

This update is a great fit for anyone working on robotics, autonomous vehicles, or vision AI who wants an accessible way to try out AI models that understand the physical world.

Related posts