NVIDIA expands Cosmos with new training tools, simulation checkpoints and data factory blueprints for world models

NVIDIA said it is expanding its Cosmos platform with new tools intended to help developers train, evaluate and deploy so-called world models, AI systems designed to understand and simulate the real world. In a new post, the company framed the update around a practical problem in physical AI: teams need large, high-quality training data and a repeatable way to test how policies and models behave before moving them into real-world conditions.

NVIDIA positioned Cosmos as modular and open, built to let developers scale training data, adapt models, test behavior, and optimize performance ahead of deployment. The announcement said the platform builds on an approach that supports multimodal, multisensory inputs as the starting point for world model development, since physical systems must often reason about more than a single data stream.

Central to the update is what NVIDIA calls Cosmos Data Factory, described as an end-to-end set of tools and blueprints for labeling, curating, augmenting and evaluating data at scale for training physical AI. NVIDIA said its GPU-accelerated solution Cosmos Curator now includes support for multimodal, multisensory inputs and additional processing blocks, aimed at improving how video datasets and related indicates are prepared for training.

The company also outlined five cloud-native blueprints included with Cosmos Data Factory and said they are available through GitHub, Microsoft Azure, Nebius and CoreWeave. NVIDIA characterized the blueprints as accelerating data curation, processing and synthetic data generation using agentic skills, with the goal of reducing friction in getting from raw footage to model-ready training sets.

NVIDIA said companies are already using the Cosmos Data Factory blueprints to speed physical AI development workflows. Porsche, for example, was described as using Cosmos Data Factory through Voxel51 on Microsoft Azure as part of a shared automated-vehicle data “flywheel” across Volkswagen Group brands. NVIDIA said the company expects the approach to reduce repeated physical test drives by up to 90% and shorten development cycle times by up to 85%.

Other examples in the update focused on synthetic data and targeted retraining. NVIDIA said ANYbotics generated more than 1,500 synthetic video clips using Cosmos Data Factory, then used Encord’s platform to validate outputs while training industrial-inspection robots to recognize liquid spills and other changes between facility visits. NVIDIA also said Ultralytics retrained its YOLO model using real images plus synthetic defects generated with Cosmos Data Factory, improving defect localization and recall.

Beyond data preparation, NVIDIA said developers can train their own architectures or start from pretrained models before deploying. At GTC Berlin, NVIDIA said it extended Cosmos Framework training, optimization and deployment tools and recipes as agent skills to more open models, including what it described as Cosmos 3. The update further said developers and agents can adapt open models such as GR00T, Alpamayo, QWEN, Pi0.5 and other compatible Hugging Face models using fine-tuning and reinforcement learning, then optimize via distillation and quantization for deployment with high-performing inference.

NVIDIA also highlighted new Cosmos 3 Nano model variants for simulation and said accompanying code enables developers to adapt the models to train and evaluate their own policies in interactive, repeatable scenarios that can be difficult to capture in real life. The company’s post also cited work by MicroAGI, which it said post-trained Cosmos 3 on videos of humans manipulating objects and distilled a 32-billion-parameter reasoner into an 8-billion-parameter Nano model without loss of accuracy, described as an effort to support egocentric manipulation research.

The company added that it is updating deployment tooling through NVIDIA Cosmos 3 NIM microservices, which it said can deliver up to 2x faster inference with no quality loss in validated configurations, alongside continuous video analysis enabling sub-second alerts. For vision AI developers, NVIDIA said it also released an NVIDIA Metropolis Blueprint for video search and summarization (VSS) 3.3 built on Cosmos NIM microservices, designed to turn one prompt into an efficient video AI application and to help teams monitor 40% more video streams with the same GPU.

Not all details were disclosed. NVIDIA did not provide independent benchmarks, performance datasets, or specific customer names tied to every claimed improvement. For some of the blueprints and deployment configurations, the post emphasizes validated configurations and expected reductions in physical testing, but it does not describe baseline measurement methods, timelines, or the scale of deployments behind those figures. Teams evaluating the platform may want to confirm how much of the benefit depends on their data pipelines, compute setup, and the types of physical tasks they are targeting.