Nvidia expanded its Cosmos platform for world models with new data, training and deployment tooling. Cosmos Data Factory now ships five cloud-native blueprints, available through GitHub and through Azure, Nebius and CoreWeave: autonomous-vehicle dataset generation with synthetic hard-case expansion, image attribute augmentation, event video generation, defect image generation for inspection models, and video augmentation that changes weather and lighting while preserving scene structure. Cosmos Curator added multimodal, multisensory inputs and processing blocks. On the training side, Cosmos Framework extends its training, optimisation and deployment recipes as agent skills to open models including GR00T, Alpamayo, Qwen and Pi0.5, alongside Cosmos 3. Nvidia cites MicroAGI distilling a 32-billion-parameter Cosmos 3 reasoner into an 8-billion-parameter Nano model with no loss in accuracy, Porsche using the blueprints through Voxel51 on Azure with expected cuts in physical test drives, ANYbotics generating more than 1,500 synthetic clips for inspection robots, and Ultralytics retraining YOLO with synthetic defect images. New Cosmos 3 Nano variants generate multiview, multisensor video for closed-loop policy evaluation, and updated NIM microservices are claimed to deliver up to 2x faster inference.
Our takeA world model is only as useful as the policy evaluation it supports, so the closed-loop Nano variants are the substantive part of this release; the data blueprints mostly industrialise curation work that teams already did by hand. Two claims deserve scrutiny. Distilling a 32B reasoner to 8B with no accuracy loss, and 2x inference with no quality loss, are both stated without benchmark detail here. Synthetic defect imagery also carries a distribution-shift risk that tends to surface on the line rather than in the eval.