The problem: latency between action and response
Kyutai has published CADRE, a family of open-source video autoencoders intended for interactive world models, that is, systems that generate video in response to an action. According to material released by the company and reported by MarkTechPost, the weak point of many world models is latency: autoencoders inherited from text-to-video generators compress multiple frames into a single latent, so each action must wait for the block to be completed. Kyutai identifies about 230 ms as the threshold beyond which a person perceives their own movement as delayed, and claims that models based on CADRE react within that limit, running at 20 fps on a single H100 GPU.
How it is built and how it was evaluated
Its operation revolves around compression, indicated in the τ×s×s format: τ is the number of frames per latent and s the spatial factor. According to the research, latency depends on τ divided by the frame rate, plus generation and decoding time. The encoder follows the MIRA recipe: a frozen DINOv3 extracts per-patch features, then a trained 3D convolution compresses them into a 64-channel latent. The decoder is a causal diffusion transformer with 24 blocks that operates on successive blocks, trained with progressive self-distillation to decode in one or two steps, in a single stage and without an adversarial loss. Training used 23 public video sources, about 76,000 hours, spanning web, robotics, driving, and video games.
For evaluation, Kyutai trained the same 1.2-billion-parameter causal transformer on each autoencoder, in four environments: robotic manipulation, human manipulation, driving, and indoor navigation. Each model observes two seconds of real video and generates 4.8 seconds. The scores cover image quality, motion quality, and ability to follow the action, measured with an inverse dynamics probe in which a value of 1 corresponds to real video.
Reported results and acknowledged limitations
The figures are self-reported by Kyutai. The CADRE-2×32×32 variant responds in 185 ms and sustains 23.4 fps with one decoding step, versus 317 ms for the Wan 2.2 autoencoder, 492 ms for LTX-2, and 605 ms for Cosmos. The gain is more pronounced with a moving camera: in driving, Wan 2.2's FVD score reaches 599.7, while CADRE-2×16×16 at one step drops to 162.6. Compared with the retrained MIRA decoder, the one-step diffusion decoder reduces FVD by a factor of 1.4 to 2.5.
The picture is not uniform, however. In robotic manipulation, Cosmos obtains an FVD of 39.5 versus 47.6 for the best CADRE variant at one step; on speed, DC-AE beats CADRE-1×48×48, despite working frame by frame. Wan 2.2 reconstructs unseen video better, with 35.02 dB PSNR. Kyutai counters that reconstruction quality does not predict rollout quality, and cites Wan 2.2 as last for FVD in indoor navigation. Seams also remain visible between blocks with long latents, and each world model covers only one environment at 288×512: transferability to general-purpose models has not been demonstrated.
Licenses and what remains to be verified
The code is distributed under the MIT license, but the weights on Hugging Face are released under the CC BY-NC 4.0 license, which excludes commercial use: a significant constraint for anyone wishing to use CADRE in products. The release includes seven autoencoders, four multi-resolution versions trained longer, and four world models based on CADRE-2×16×16, plus an interactive demo controllable by keyboard. Quantized variants are not specified. Before drawing conclusions about real-world performance, independent verification must be awaited: at present, the available data come from the same research that introduces the model.
