NVIDIA Cosmos Predict 2.5 and Transfer 2.5

2025-10-28 · Publisher : NVIDIA · Domain : Robotics

NVIDIA releases Cosmos Predict 2.5, merging three models into a single architecture trained on 200 million clips, and Cosmos Transfer 2.5, faster and 3.5x smaller than its predecessor.

### Background On October 28, 2025, NVIDIA released Cosmos Predict 2.5 and Cosmos Transfer 2.5, new versions of its open world models announced a month earlier at CoRL. Weights are on Hugging Face, with a technical post detailing the changes. ### Specifications - **Cosmos Predict 2.5**: merges three previously separate models, Text2World, Image2World and Video2World, into a single architecture generating consistent, controllable video worlds from several input types. - **Data**: trained on 200 million high-quality video clips. - **Capabilities**: videos up to 30 seconds and multi-view camera outputs for richer simulations, per NVIDIA. - **Cosmos Transfer 2.5**: photoreal synthetic data from simulated 3D scenes or spatial inputs (depth, segmentation, edges), faster, higher quality and 3.5x smaller than the previous version. - **License**: NVIDIA open model license. ### Why it matters Longer, multi-view videos bring world models closer to real needs: a robot or vehicle perceives the scene through several cameras over long sequences. Merging into one model also simplifies developer use. ### Key takeaway Cosmos 2.5 became the synthetic data base of NVIDIA's stack for 2026, ahead of Cosmos 3 announced at GTC 2026.

Tags : Modèle du monde, NVIDIA, Données synthétiques, Poids ouverts