Kyutai releases CADRE video autoencoders for interactive world models

Kyutai’s blog post announcing CADRE is dated 22 October 2026.[2] The release includes open models at seven spatiotemporal compression settings.[3] The autoencoders differ in a small bottleneck applied to frozen features and share a diffusion decoder.[3] More specifically, the encoder uses a frozen DINOv3 feature extractor and a trained 3D convolution to compress video into a latent representation.[1] The decoder is self-distilled to run in one or a few steps.[3] Kyutai says the autoencoders were not designed with the throughput-latency trade-off of world models in mind.[3]

Kyutai trained the autoencoders on about 76,000 hours of video from 23 public sources, spanning web, robot, driving and game footage.[1] The study then trained the same 1.2-billion-parameter world-model recipe on CADRE autoencoders and compared results across four environments.[4] Those environments cover robot manipulation, human manipulation, driving and indoor navigation.[1]

In the reported interactive operating point, CADRE-2×32×32 with a one-step decoder reached 185 milliseconds of action latency and a maximum 23.4 frames per second on one H100.[4] The paper also describes its featured setting as running at 20 frames per second on one H100 and reacting in under 230 milliseconds, a threshold it identifies as where people notice lag in their own movement.[3] At one H100, reported action latency was 317 milliseconds for Wan 2.2, 492 milliseconds for LTX-2 and 605 milliseconds for Cosmos.[1] One-step CADRE had 1.4-to-2.5-times lower FVD across the four environments.[4]

Compression choices matter: moving from spatial compression 16× to 32× raised FVD@96 from 39.4 to 95.2 and lowered ARR@96 from 0.73 to 0.45, while FID barely moved.[4] The paper says temporal compression costs little when the camera is still, while spatial compression matters when it moves.[3] CADRE did not lead every comparison: Cosmos recorded a robot-manipulation FVD of 39.5, below 47.6 for the best one-step CADRE result.[1] The study says whether its findings transfer to generalist world models remains open.[4] Hugging Face recommends CADRE-2×16×16 as a starting autoencoder, and the weights’ CC BY-NC 4.0 license excludes commercial use.[5][1]