Kyutai Releases CADRE Video Autoencoders for World Models With Sub-230 ms Latency

The release contains seven autoencoders, each with about 476 million trained parameters plus a frozen 86 million-parameter DINOv3 ViT-B/16. The world model used in the experiments is a 1.2 billion-parameter causal diffusion transformer. Video runs at 288×512 and 20 fps on one H100 in bf16 with batch size 1. Kyutai did not disclose quantized variants. The weights are licensed under CC BY-NC 4.0 plus the DINOv3 License.

CADRE is built around compression. The autoencoder writes compression as τ×s×s, where τ is frames per latent and s is the spatial factor. Because the world model generates in the autoencoder’s latent space, the compression sets both cost per frame and action delay. Latency is τ/f plus generation and decode time. At τ = 8 and 20 fps, the chunk alone adds 400 ms, while humans notice their own movement lagging beyond about 230 ms. The encoder reuses the MIRA recipe: a frozen DINOv3 extracts per-patch features from each frame, and one trained 3D convolution compresses them into a 64-channel latent. Only this bottleneck changes across the seven members. The decoder is a causal diffusion transformer with 24 blocks that streams chunk by chunk; progressive self-distillation teaches it to decode in one or two steps in a single training stage without an adversarial loss. Training used 23 public video sources totaling about 76,000 hours across web, robot, driving and game video.

Kyutai trained the same 1.2 billion-parameter causal diffusion transformer on each autoencoder and tested it in four environments: robot manipulation (ABC), human manipulation (EgoDex), driving (PhysicalAI AV) and indoor navigation (RealEstate10K). Each model sees 2 seconds of real video and generates 4.8 seconds. Scores cover image quality (FID), motion quality (FVD) and action following (ARR), with ARR using a frozen inverse dynamics probe; an ARR of 1 matches real video.

On one H100, CADRE-2×32×32 reacts in 185 ms and sustains 23.4 fps with one decoding step. The Wan 2.2 autoencoder needs 317 ms, LTX-2 492 ms and Cosmos 605 ms. DC-AE is faster at 118 ms but operates frame by frame. The gap is widest with a moving camera. In driving, Wan 2.2 reaches an FVD of 599.7 and LTX-2 913.3, while CADRE-2×16×16 with one step posts 162.6. In indoor navigation, LTX-2’s ARR falls to 0.04. Against the retrained MIRA decoder, the one-step diffusion decoder lowers FVD by 1.4 to 2.5 times. Kyutai reports that CADRE forms most of the latency-quality frontier and is as fast and better in 42 of 48 comparisons with MIRA, Wan 2.2, Cosmos and LTX-2.

The paper lists four findings. Decoding by diffusion avoids the blur of perceptual regression at strong compression, while diffusion samples one sharp video. Compressing time inside the autoencoder preserves detail that patching inside the world model loses; at eight frames, ARR drops from 0.72 to 0.56. Camera motion decides compression: still cameras tolerate temporal compression, while moving cameras favor fewer frames and coarser spatial tokens. Coarse tokens hurt motion first: in robot manipulation, 32× spatial compression multiplies FVD by 2.4 over 16×, while FID barely moves.

CADRE does not win everywhere. In robot manipulation, its ARR falls to 0.45 for CADRE-2×32×32 against 0.85 for Wan 2.2, and Cosmos posts an FVD of 39.5 against 47.6 for the best one-step CADRE. DC-AE beats CADRE-1×48×48 in manipulation. Wan 2.2 reconstructs held-out video far better, at 35.02 dB PSNR. Kyutai argues reconstruction does not predict rollouts; Wan 2.2 ranks last on indoor rollout FVD. Long latents leave visible seams between chunks, and at τ = 8 boundary excess reaches 1.6 to 1.8 across environments.