Kyutai CADRE: Video Autoencoders That Cut World Model Lag Below 230 ms

Why it matters

Kyutai has released CADRE, a family of open video autoencoders designed specifically for action-conditioned world models rather than borrowed from text-to-video systems. The core problem it addresses is action lag: conventional autoencoders compress many frames into a single latent, forcing every user action to wait for a chunk to finish. CADRE instead compresses time inside the autoencoder and uses a causal diffusion decoder that streams chunk by chunk, learning to decode in very few steps through progressive self-distillation without adversarial training. The release includes multiple autoencoders that differ only in their compression bottleneck, letting Kyutai isolate compression as the single variable when training identical world models across robot manipulation, human manipulation, driving and indoor navigation environments. The benchmarks, all self-reported, show CADRE forming most of the latency-quality frontier against MIRA, Wan 2.2, Cosmos and LTX-2, though it loses in some robot manipulation and reconstruction measures. Code is MIT licensed while weights carry a non-commercial license, and a live keyboard-driven demo accompanies the release.