Research & Innovation
Kyutai Rebuilt the Video Autoencoder Because World Models Were Too Slow to Feel Like Anything
Most world models borrow an autoencoder from a text-to-video generator, which packs many frames into one latent. That is fine for generating a clip and fatal for anything interactive: every action waits for the chunk. Kyutai's CADRE puts a trained bottleneck on frozen DINOv3 features and decodes with a causal diffusion transformer in one or two steps. Code is MIT, weights are published, and there is a demo you can drive in a browser — but the paper is still 'coming soon'.