Kyutai has released CADRE, a family of 7 open causal diffusion video autoencoders built for interactive world models — and the headline number is under 230 milliseconds of action lag running at 20 frames per second on a single NVIDIA H100.
Kyutai CADRE (Credits: MarkTechPost)
World models — AI systems that simulate environments so robots and agents can learn inside them — have a lag problem. Many borrow their video autoencoders from text-to-video generators, which pack many frames into a single latent. That compression is great for generating clips, but it means every action has to wait. Humans notice movement lagging beyond about 230 ms, and CADRE is engineered to stay under that line.
How Kyutai CADRE Cuts World Model Lag
CADRE works directly in the world model’s latent space. Each autoencoder has about 476M trained parameters plus a frozen 86M DINOv3 ViT-B/16, with a 1.2B-parameter world model generating video at 288×512 and 20 fps. The decoder is a causal diffusion transformer with 24 blocks that streams chunk by chunk, using progressive self-distillation to decode in just 1 or 2 steps.
The results are striking. CADRE-2×32×32 reacts in 185 ms and sustains 23.4 fps on one H100 — against 317 ms for the Wan 2.2 autoencoder, 492 ms for LTX-2 and 605 ms for Cosmos. It also wins on quality: CADRE was faster and better in 42 of 48 comparisons against MIRA, Wan 2.2, Cosmos and LTX-2, including a driving FVD score of 162.6 versus 599.7 for Wan 2.2.
Trained on 76,000 Hours of Video
The autoencoders trained on 23 public video sources — about 76,000 hours spanning web footage, robot video, driving data and game video. Kyutai evaluated the same 1.2B model on four environments: robot manipulation, human manipulation, driving and indoor navigation, where each model watches 2 seconds of real video then generates 4.8 seconds.
The honest weak spot: robot manipulation action following, where CADRE’s ARR score of 0.45 trails Wan 2.2’s 0.85. But for interactive simulation — the latency-sensitive use case CADRE was built for — it now defines the latency-quality frontier.
Also Read: You.com Launches Inference Router: 12 Open-Weight Models With 99.9% Uptime SLA | GM Targets Eyes-Off Autonomous Driving for Pickups, SUVs to Chase Software Revenue
Source: MarkTechPost — read the full story there.
