World foundation model platform for physical AI. Cosmos-Predict2.5 uses flow-based architecture for Text2World/Image2World/Video2World at 2B and 14B scales. Trained on 200M curated video clips with RL-based post-training. One of NVIDIA's six frontier model families alongside Nemotron, GR00T, Alpamayo, BioNeMo, and Earth-2.

Paper

arXiv: 2501.03575

generationvideoopen-weight