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

NVIDIA has expanded its Cosmos platform for building world models, the systems that help physical AI understand and simulate real environments. In an October 21 blog post tied to GTC Berlin, the company announced new training tools, simulation checkpoints and Cosmos Data Factory blueprints. Five cloud-native blueprints, available through GitHub, Microsoft Azure, Nebius and CoreWeave, cover autonomous vehicle datasets, image attribute augmentation, event video generation, defect imagery and video augmentation.

NVIDIA cites several early adopters. Porsche, working through Voxel51 on Azure, is leading a shared autonomous driving data effort across Volkswagen Group brands that is expected to cut repeated physical test drives by up to 90% and development cycles by up to 85%. ANYbotics generated more than 1,500 synthetic clips to train inspection robots, while Ultralytics says retraining YOLO with targeted synthetic defects improved localization and recall.

Cosmos Framework tools now extend, as agent skills, to open models including GR00T, Alpamayo, QWEN and Pi0.5, supporting fine-tuning, reinforcement learning, distillation and quantization. According to NVIDIA, research lab MicroAGI distilled Cosmos 3's 32-billion-parameter reasoner into an 8-billion-parameter Nano model with no accuracy loss.

New Cosmos 3 Nano variants let developers test autonomous vehicle and two-armed robot policies in simulation before real-world trials. NVIDIA also claims updated NIM microservices deliver up to twice the inference speed in validated configurations, and that its video search blueprint monitors 40% more streams per GPU. ADNOC uses a Fogsphere platform built on these tools to spot safety hazards across 65 sites. The update matters because it targets physical AI's costly bottlenecks: scarce data and real-world testing.