Cosmos 3

NVIDIA Cosmos 3 Physical AI flywheel — omnimodal world models for generate → post-train → eval

Overview

Cosmos 3 on AWS is a runnable sample for the NVIDIA Cosmos 3 Physical AI flywheel — a family of omnimodal world models for Physical AI, openly licensed under OpenMDW-1.1. The same model family serves as both the synthetic-data generator and the policy you post-train.

This sample post-trains an action policy on turnkey public LeRobot v3 datasets (e.g. lerobot/droid_100, BridgeData2, LIBERO) using each dataset’s native action vectors, deployed on Amazon EKS / SageMaker HyperPod.

Key Features

• Full flywheel: generate → post-train → eval pipeline
• Omnimodal world models — vision + action in a single architecture
• Validated end-to-end on p5en.48xlarge (8× H200) with real checkpoints
• Multiple training modes: SDG, action-policy post-training, Nano vision SFT, Super-64B LoRA
• EKS + HyperPod deployment (same control plane)