NVIDIA expands Cosmos tools for training and deploying world models

NVIDIA expanded its Cosmos world-model platform on Wednesday with training and optimization tools, Data Factory blueprints, simulation checkpoints and deployment options for developers building robotics, autonomous-vehicle and video-analysis systems.

Together, the additions broaden the public development stack, from data preparation and labeling to model adaptation, repeatable simulation and deployment. Code is available in public repositories, simulation variants are listed on Hugging Face, and NVIDIA has published a Cosmos 3 NIM container through its NGC catalog.

The Cosmos Data Factory repository lists five reference blueprints: autonomous-vehicle dataset generation, defect-image generation, event-video generation, image-attribute augmentation and video-data augmentation. Its composable skills cover curation and retrieval, augmentation, auto-labeling, orchestration, simulation and synthetic-data evaluation. The repository says the workflows can run on premises or in a chosen cloud. Its setup requirements include Kubernetes, object storage, a writable container registry and access to NVIDIA NGC and Hugging Face.

The public Cosmos Framework provides components for training, inference, policy serving, post-training, reinforcement learning and distillation. NVIDIA’s documentation lists integrations for Cosmos 3, GR00T-N1.7, pi0.5, Alpamayo through AlpaGym, Qwen model families and compatible Hugging Face backends.

NVIDIA also released 15-billion-parameter Cosmos 3 Nano simulation variants for autonomous-driving and bimanual-robot scenarios, plus a Sim-Transfer variant, through its Cosmos 3 examples collection. The company says accompanying code lets developers adapt the models for repeatable multiview and multisensor simulations before real-world testing. DataPhoenix has separately examined NVIDIA’s Cosmos3-H-Dreams release for surgical-simulation research, another use of Cosmos as a simulation system. Neither release establishes real-world safety or eliminates the need for physical testing.

For deployment, NVIDIA made an updated Cosmos 3 NIM container available through NGC. The company claims the updated NIM microservices can run inference up to twice as fast without losing quality. Its announcement, however, does not disclose the baseline, hardware, workload or complete test results needed to verify that figure.

NVIDIA separately claims its Metropolis Video Search and Summarization Blueprint 3.3 can monitor 40% more video streams on the same GPU. The accessible VSS product page confirms that the blueprint can be deployed, but it does not provide the versioned benchmark methodology.

The announcement also attributes several results and forecasts to Cosmos users. NVIDIA says ANYbotics generated more than 1,500 synthetic clips, MicroAGI distilled a 32-billion-parameter reasoner into an 8-billion-parameter model without losing accuracy, and ADNOC uses a Cosmos-based platform at 65 sites. The company also says a planned Porsche workflow could reduce repeated test drives by up to 90% and development cycles by up to 85%. The reviewed sources do not independently substantiate those figures.

NVIDIA says Cosmos Data Factory is available through GitHub, Microsoft Azure, Nebius and CoreWeave. GitHub availability is directly observable. Microsoft describes a planned autonomous-vehicle workflow using Cosmos Data Factory on Azure. Meanwhile, CoreWeave’s launch catalog includes Data Factory skills for augmentation, labeling, curation and retrieval, with additional workflows described as coming soon. The reviewed evidence did not independently confirm the announced Nebius availability or establish uniform service regions, pricing or access terms across the channels.