NVIDIA has teamed up with researchers from MIT and the University of Oxford to unveil the Physis-Lang solution. This innovative approach seamlessly incorporates physical causes, laws, and their effects into language descriptions, applying them comprehensively throughout the processes of data filtering, model training, and video generation. To rigorously assess the adequacy and accuracy of physical details within video descriptions, the team has developed the PhysCapBench evaluation system. This system fosters the self-evolution of description criteria by fixing annotation model weights and iteratively refining shared caption prompts. Moreover, the Physis-Lang solution can pinpoint the model's physical shortcomings and transform them into physical domain labels, enabling the targeted retrieval and supplementation of training data. The Cosmos3-Super and Cosmos3-Nano models, enhanced with the Physis-Lang solution, clinched the top two spots on the Physics-IQ Verified leaderboard, significantly outperforming the previous best scores. Simultaneously, they achieved open-source SOTA (State-of-the-Art) results across four physical evaluations, surpassing Veo 3.1 in three of these assessments and generating video content that more closely adheres to real-world physical laws.