Cosmos3-Super-Reasoner

Cosmos3-Super-Reasoner is a fully open 64B-parameter omnimodel (a 32B reasoner paired with a 32B generator) built to advance physical AI research through native multimodal reasoning and generation. It is designed for large-scale synthetic data generation and policy model development, natively understanding and generating text, images, video, ambient sound, and actions in unified workflows. The model targets robotics, autonomous systems, and vision AI agents, combining world generation, physical reasoning, and action generation in a single model rather than stitching together several specialized ones. NVIDIA positions it as one of the first fully open omnimodels of its kind for physical AI. With a 262K token context, it handles long-horizon reasoning scenarios and is suited to researchers and engineers building foundation models for physical systems, training robotic policies from synthetic data, and developing agentic AI for embodied task planning.