Kyutai releases CADRE video autoencoders for world models with sub-230 ms lag

Kyutai has open-sourced CADRE, a family of causal diffusion video autoencoders for interactive world models. On a single H100 the models run at 20 fps with reaction times under 230 ms and beat MIRA, Wan 2.2, Cosmos and LTX-2 in 42 of 48 comparisons. Code is MIT-licensed, while Hugging Face weights are CC BY-NC 4.0, excluding commercial use.

• 7 autoencoders of about 476M parameters each plus a frozen 86M DINOv3 ViT-B/16
• 1.2B-parameter world model, 288×512 video at 20 fps on one H100
• Driving FVD of 162.6 versus 599.7 for Wan 2.2; robot ARR 0.45 versus 0.85
• Trained on 23 public video sources, roughly 76k hours