NVIDIA’s Cosmos Reason 2 is an open reasoning vision‑language model that surpasses its predecessor in accuracy and leads the Physical AI Bench and Physical Reasoning leaderboards as the top open model for visual understanding. The new release offers 2B and 8B parameter variants, a 256K token context window, and expanded spatial and visual perception features including OCR, 2D/3D point localization, bounding box coordinates, trajectory data, and improved timestamp precision. Cosmos Reason 2 is optimized for flexible deployment from edge to cloud and is integrated into a range of use cases such as video analytics, robot planning, and autonomous vehicle training. Companies such as Uber, Salesforce, Encord, Hitachi, Milestone, and VAST Data are leveraging the model for tasks like video captioning, VQA, and safety monitoring. The model is available on Hugging Face and will soon be offered on AWS, GCP, and Azure, with documentation and a Cosmos Cookbook to facilitate development. The release is part of NVIDIA’s broader Cosmos ecosystem, which includes generative models like Cosmos Predict and Cosmos Transfer, and the humanoid‑robot‑focused VLA model NVIDIA GR00T N1.6.
NVIDIA Cosmos Reason 2 Brings Advanced Reasoning To Physical AI
NVIDIA has released Cosmos Reason 2, an open reasoning vision‑language model that tops Physical AI Bench and Physical Reasoning leaderboards. The model offers 2B and 8B parameter versions, 256K token context, OCR, 2D/3D point localization, trajectory data, and improved spatio‑temporal understanding. It supports edge to cloud deployment and is integrated into video analytics, robot planning, and autonomous vehicle training workflows.
Why it matters
The model’s advanced reasoning and spatio‑temporal capabilities enable more robust physical AI applications, from autonomous driving to workplace safety, potentially accelerating deployment of AI in real‑world environments.