Cosmos Reason

Cosmos Reason is NVIDIA's open reasoning vision-language model for physical AI and robotics, first unveiled at GTC 2025. The 7-billion-parameter model takes video and text, reasons step by step using prior knowledge, physics understanding, and common sense, and produces detailed logical responses for robot planning, data curation, and video analytics.

It is built on a Qwen2.5-VL-7B-Instruct backbone with supervised fine-tuning plus reinforcement learning, using chain-of-thought reasoning without human annotations. NVIDIA reports a 65.7 average score across robotics and autonomous vehicle benchmarks and top placement on the Physical Reasoning leaderboard. The weights are openly available on Hugging Face under NVIDIA's open model license, with fine-tuning guides for robotics-specific tasks. It anchors NVIDIA's Cosmos world-foundation-model platform alongside the generative Cosmos world models.