Cosmos Data Factory now has five cloud-native blueprints covering autonomous-vehicle dataset generation, image attributes, event-video generation, defect-image generation and video augmentation. NVIDIA says they are available through GitHub, Microsoft Azure, Nebius and CoreWeave.
The company also extended Cosmos Framework training, optimization and deployment tools and recipes as agent skills for open models including GR00T, Alpamayo, Qwen, Pi0.5 and compatible Hugging Face models. The framework supports fine-tuning, reinforcement learning, distillation, quantization and inference deployment. NVIDIA introduced Cosmos 3 Nano simulation variants and accompanying code for adapting models to train and evaluate autonomous-vehicle and bimanual-robot policies in interactive simulations.
Updated Cosmos 3 NIM microservices are claimed to provide up to 2x faster inference with no quality loss in validated configurations. NVIDIA also cites deployments and performance claims from Porsche, ANYbotics, Ultralytics and ADNOC. The source does not establish independent benchmarks, general availability terms for every component, or whether the reported customer results will transfer to other workloads.