NVIDIA, MIT and Oxford unveil Physis-Lang for physics-accurate video

Researchers from NVIDIA, MIT and Oxford introduced Physis-Lang, a text-based representation of physics for training video world models. On the September 29, 2026 Physics-IQ Verified leaderboard, Cosmos3-Super with Physis-Lang ranked first at 48.2 ± 1.4, while Cosmos3-Nano took second at 43.3 ± 1.5. No code or weights have been released yet.

• Cosmos3-Super with Physis-Lang scores 48.2 ± 1.4 on Physics-IQ Verified, ranking first
• Cosmos3-Nano with Physis-Lang beats Veo 3.1 on 3 of 4 benchmarks
• Caption F1 rose from 78.64 at iteration 1 to 87.82 at iteration 9
• Training set has 183K videos: 71K filtered from WISA-80K plus 112K retrieved clips