NVIDIA has released the weights for its Cosmos-Reason2-32B model, a physics-aware AI reasoning vision-language model designed to enhance robotic and autonomous driving systems. Initially launched with smaller versions, the flagship 32-billion-parameter model is now available for commercial use under the NVIDIA Open Model License. Built on the Qwen3-VL-32B-Instruct foundation, the model can analyze driving videos in real time and label 2D/3D coordinates in warehouse photos. The Cosmos-Reason2-32B model is tailored for applications such as analyzing video streams from urban and industrial settings, annotating sensor data, and serving as a planning brain for robots and autonomous vehicles. It features improved object detection, precise timestamp localization, and an expanded context window of 256K tokens, offering significant advancements over its predecessors.