NVIDIA Cosmos Opens New Tools to Train, Evaluate and Deploy World Models
nvidia ·
NVIDIA has expanded its Cosmos platform with new tools designed to accelerate the training, evaluation, and deployment of world models for physical AI systems. The update introduces Cosmos Data Factory blueprints that assist in curating, labeling, and augmenting multimodal data at scale, featuring applications for autonomous vehicles and industrial inspection. Companies such as Porsche and ANYbotics are already utilizing these tools to reduce physical testing requirements and train robots on synthetic anomaly data. Developers can now adapt open models like GR00T and QWEN using the Cosmos Framework, which supports fine-tuning, distillation, and quantization to optimize performance. A new Cosmos 3 Nano model allows researchers to distill large reasoners into smaller variants without losing accuracy, facilitating faster deployment. Additionally, updated NIM microservices provide up to two times faster inference speeds, enhancing real-time video analysis capabilities. These advancements enable organizations like ADNOC to detect safety hazards more efficiently by combining synthetic data generation with agentic platforms.
nvidia