NVIDIA Expands Cosmos Tools to Train, Evaluate and Deploy World Models
NVIDIA has expanded its Cosmos platform with new tools, simulation checkpoints and cloud blueprints to help developers train, evaluate and deploy world models for physical AI. The company said Cosmos ranks among the top five open models on six leaderboards and offers frontier open models together with data, training, optimization and inference recipes for scaling robotics and autonomous systems. The refreshed Cosmos Data Factory provides cloud-native blueprints for labeling, curating, augmenting and generating multimodal, multisensory training data at scale. Cosmos Curator has added support for broader video and sensor processing workflows, and the blueprints are available through GitHub, Microsoft Azure, Nebius and CoreWeave. Porsche is using the platform with Voxel51 on Microsoft Azure to build a shared autonomous-vehicle data flywheel across Volkswagen Group brands, which NVIDIA expects could reduce repeated physical test drives by up to 90% and shorten development cycles by up to 85%. ANYbotics generated more than 1,500 synthetic video clips and validated outputs with Encord to train industrial inspection robots to detect spills and facility changes, while Ultralytics retrained its YOLO model using real images and targeted synthetic defects to improve localization and recall. At GTC Berlin, NVIDIA extended Cosmos Framework training, optimization and deployment capabilities as agent skills across additional open models, including GR00T, Alpamayo, Qwen, Pi0.5 and compatible Hugging Face models. Developers can fine-tune or reinforce model behavior, apply distillation and quantization, and deploy high-performance inference. Cosmos 3 also supports specialized post-training, as shown by robotics lab MicroAGI, which post-trained the model on human manipulation videos and distilled a 32-billion-parameter reasoner into an 8-billion-parameter Nano model without accuracy loss. Developers can also use Roboflow to label data, post-train Cosmos 3 for specific environments and deploy models on hosted or on-premises NVIDIA GPUs. For testing and deployment, NVIDIA introduced Cosmos 3 Nano simulation variants for repeatable, hard-to-capture multiview and multisensor evaluation of autonomous vehicle and bimanual robot policies. Updated Cosmos 3 NIM microservices deliver up to twice the inference speed in validated configurations while supporting continuous video analysis with sub-second alerts. The Metropolis Blueprint for video search and summarization 3.3, built on Cosmos NIM, enables teams to create video AI applications from a single prompt and monitor 40% more video streams with the same GPU capacity. ADNOC is deploying Fogsphere MistIQ, built with Cosmos and Metropolis, to detect safety hazards across 65 sites, using Cosmos Data Factory synthetic data to improve rare-event detection without staging dangerous scenarios. NVIDIA said the platform remains modular and open, letting developers generate data, train models and deploy world models through NIM microservices.
