News
NVIDIA added Cosmos training tools, simulation checkpoints and Data Factory blueprints for world models used in physical AI.
NVIDIA said it has expanded its Cosmos platform with new training tools, simulation checkpoints and Cosmos Data Factory blueprints to help developers build world models for physical AI systems. The announcement, published by NVIDIA on Oct. 22, 2026, adds to a platform the company says ranks in the top five on six open-model leaderboards and is aimed at moving projects from data preparation to deployment.
Evidence and context
According to NVIDIA, Cosmos Data Factory brings together tools for labeling, curating, augmenting and evaluating multimodal training data at scale. The company said Cosmos Curator now supports multimodal, multisensory inputs and processing blocks. Cosmos Data Factory is available through GitHub, Microsoft Azure, Nebius and CoreWeave, and includes cloud-native blueprints for autonomous-vehicle dataset generation, image attribute augmentation, event video generation, defect image generation and video data augmentation.
NVIDIA described several early uses of the blueprints. It said Porsche is using Cosmos Data Factory through Voxel51 on Microsoft Azure to build a shared autonomous-vehicle data flywheel across Volkswagen Group brands, with expected reductions of up to 90% in repeated physical test drives and up to 85% in development cycle time. The company also said ANYbotics generated more than 1,500 synthetic video clips with Cosmos Data Factory blueprints to train industrial-inspection robots, while Ultralytics retrained its YOLO model using real images and synthetic defect data to improve defect localization and recall.
The updated Cosmos Framework extends training, optimization and deployment tools and recipes as agent skills to more open models, NVIDIA said. The company said developers can adapt open models including GR00T, Alpamayo, QWEN and Pi0.5, as well as compatible Hugging Face models, through fine-tuning and reinforcement learning. It also said models can be optimized with distillation and quantization before deployment using high-performing inference.
NVIDIA said MicroAGI post-trained Cosmos 3 on human-manipulation videos and distilled its 32-billion-parameter reasoner to an 8-billion-parameter Nano model with no loss in accuracy. The company also said developers can label data and post-train Cosmos 3 on Roboflow, then deploy models on Roboflow hosted or on-premises NVIDIA GPUs for video reasoning in factories and warehouses.
On simulation, NVIDIA said new Cosmos 3 Nano variants are designed to test autonomous policies in repeatable multiview, multisensor scenarios, with code for adapting the models to train and evaluate policies for autonomous vehicles and bimanual robots in interactive simulations before real-world testing. The company also said updated Cosmos 3 NIM microservices can deliver up to 2x faster inference with no quality loss in validated configurations, while continuous video analysis can enable sub-second alerts.
For video applications, NVIDIA said its new Metropolis Blueprint for video search and summarization 3.3 is built on Cosmos NIM microservices and can help teams monitor 40% more video streams with the same GPU. The company said ADNOC uses Fogsphere’s MistIQ agentic platform, built with Cosmos and the Metropolis VSS Blueprint, to detect safety hazards across 65 sites, and that Fogsphere uses Cosmos Data Factory to generate synthetic training data for rare events.
Background material supplied with the announcement says NVIDIA launched Cosmos on Jan. 6, 2025 and Cosmos 3 on June 1, 2026. NVIDIA’s repository also lists an October 2026 interactive-world-simulation update with example checkpoints and training recipes, but no exact day was provided in the supplied material, so it is referenced only as an October update.
Be the first to comment