A world model that responds half a second late is not a simulator, it is a slideshow. Kyutai’s new CADRE family attacks that failure. The release — Rethinking Video Autoencoders for Interactive World Models, on arXiv since 21 October — bundles seven open causal diffusion representation autoencoders whose only structural difference is a small bottleneck on frozen features. That design is the point: it makes compression the single controlled variable, so the latency-quality frontier is measured rather than asserted.
Why Autoencoders Set the Lag
A world model generates inside its video autoencoder’s latent space, so the autoencoder’s compression sets both cost per frame and how long each action waits to take effect. Most world models borrow autoencoders from text-to-video generators, which pack many frames into one latent — fine for offline generation, but every action queues behind the chunk. At 8 frames per latent and 20 fps, the chunk alone adds 400 ms. Humans notice their own movement lagging past roughly 230 ms.
CADRE’s encoder reuses the MIRA recipe: a frozen DINOv3 ViT-B/16 extracts per-patch features from each frame, and one trained 3D convolution compresses them into a 64-channel latent. The decoder is a causal diffusion transformer with 24 blocks that streams chunk by chunk, self-distilled to decode in one or two steps — one training stage, no adversarial loss. Each autoencoder carries roughly 476M trained parameters on top of the frozen 86M backbone, trained on 23 public video sources totalling about 76,000 hours.
The Numbers on One H100
The world model is 1.2B parameters, runs at 288×512 and 20 fps on a single H100 in bf16. On CADRE-2×32×32 reaction time is 185 ms at 23.4 fps with one decoding step; Wan 2.2 needs 317 ms, LTX-2 492 ms, Cosmos 605 ms. Only DC-AE is faster at 118 ms, but decodes frame by frame.
Quality gaps widen with a moving camera. On driving, CADRE-2×16×16 posts an FVD of 162.6 against 599.7 for Wan 2.2 and 913.3 for LTX-2. On robot manipulation, action-following rate reaches 0.73 for CADRE-2×16×16 versus 0.85 for Wan 2.2 and 0.25 for LTX-2.
Four findings come out of it: diffusion decoding beats perceptual regression under strong compression; compressing time inside the autoencoder beats patching inside the world model (at 8 frames, action-following drops from 0.72 to 0.56); camera motion decides the trade — still cameras tolerate temporal compression, moving cameras prefer fewer, coarser frames; and coarse spatial tokens hurt motion before image quality, with 32× spatial compression multiplying FVD by 2.4 over 16× while FID barely moves.
Where It Loses
CADRE does not win everywhere. Cosmos posts a better manipulation FVD (39.5 against 47.6), Wan 2.2 reconstructs held-out video better at 35.02 dB PSNR, and long latents leave visible seams. CADRE-4×24×24 is an outlier at FVD 403 on driving, and each world model covers one environment, so transfer to generalist models is open. Licensing deserves a note too: code is MIT, but weights are CC BY-NC 4.0, excluding commercial use.
The release ships 7 autoencoders and 4 world models on CADRE-2×16×16, with a keyboard-driven demo. The most quotable number is not an FVD: action lag now fits inside the human perception budget on one GPU. Related: Rethinking Video Autoencoders for Interactive World Models (arXiv:2610.30129)