Kyutai Releases CADRE Video Autoencoders

Kyutai released CADRE, open causal diffusion representation autoencoders for interactive world models. The system runs at 20 fps on 1 H100 and reacts within 230 ms. It features 7 autoencoders with 476M trained parameters each plus a frozen 86M DINOv3 ViT-B/16 model.