Build a Physical AI model factory with NVIDIA Cosmos 3 on SageMaker HyperPod
2026-09-05 00:16Models🔥 42.2 heat score
1sources
1days unfolding
42.2heat score
4mentions
SummaryAI generated
On September 4, 2026, NVIDIA announced on its machine learning blog the launch of Cosmos 3, designed to utilize Amazon SageMaker HyperPod to build a physical AI model factory. This factory generates synthetic data through continuous loops, trains and推理 strategy models, and evaluates them in a closed-loop simulation, transforming real-world data into physical actions. Cosmos 3 adopts a hybrid Transformer (MoT) architecture with joint attention at each layer and an asymmetric design for training and inference. As an open multi-modal world base model, it treats video, images, actions, and sounds as a single token stream and operates the same backbone network in three modes: the forward dynamic world model is used for generating synthetic videos, the reverse dynamic action labeler for post-training, and the deployable action strategy for evaluation. This design eliminates the need to configure GPU resources separately for each stage in traditional physical AI pipelines, scheduling generation, training, and evaluation workloads on the same persistent and highly available G…
NVIDIA released Cosmos 3, designed to build a physical AI model factory based on Amazon SageMaker HyperPod. This factory generates synthetic data through continuous loops, trains and推理 perceptual and strategic models, and evaluates them in closed-loop simulations, transforming real-world data into physical actions. Cosmos 3 adopts a hybrid Transformer (MoT) architecture with a combined attention mechanism at each layer and an asymmetric design for training and inference. As an open multimodal world base model, it treats video, images, actions, and sounds as a single token stream and operates the same backbone network in three modes: the forward dynamic world model is used for generating synthetic videos, the reverse dynamic action tagger for post-training, and the deployable action strategies for evaluation. This design eliminates the need to configure GPU resources separately for each stage in traditional physical AI pipelines, scheduling generation, training, and evaluation workloads on the same persistent and highly available GPU node pool for time sharing. The model is based on Linux F…