NVIDIA Cosmos Opens New Tools to Train, Evaluate and Deploy World Models

NVIDIA’s Cosmos platform now includes training tools, simulation checkpoints, and Cosmos Data Factory blueprints to help developers build world models for physical AI systems. These models enable systems to understand and simulate real-world environments. The platform provides frontier open models, data tools, and optimization recipes to scale training, adapt models, and test performance before deployment. Developers can access these resources through GitHub, Microsoft Azure, Nebius, and CoreWeave.

Cosmos Data Factory offers cloud-native blueprints for data curation, processing, and synthetic data generation, supporting multimodal and multisensory inputs. Companies like Porsche and ANYbotics are using these blueprints to accelerate physical AI development. Porsche’s autonomous vehicle data flywheel aims to reduce test drives by up to 90% and shorten development cycles by up to 85%. ANYbotics generated over 1,500 synthetic video clips to train industrial inspection robots.

At GTC Berlin, NVIDIA introduced Cosmos Framework tools and Cosmos 3 to streamline training, optimization, and deployment of world models. Developers can fine-tune open models like GR00T and QWEN, optimize with distillation and quantization, and deploy with high-performing inference. MicroAGI post-trained Cosmos 3 on human manipulation videos, distilling a 32-billion-parameter model to an 8-billion-parameter Nano model without accuracy loss.

NVIDIA also released Cosmos 3 Nano model variants for simulation, enabling developers to test autonomous policies in repeatable scenarios. Updated Cosmos 3 NIM microservices deliver up to 2x faster inference with no quality loss, while the NVIDIA Metropolis Blueprint for video search and summarization 3.3 helps monitor 40% more video streams with the same GPU. ADNOC uses Fogsphere’s MistIQ platform, built with Cosmos and the Metropolis VSS Blueprint, to detect safety hazards in real time across 65 sites.