Kyutai CADRE: Video Autoencoders That Cut World Model Lag Below 230 ms

Kyutai has released CADRE, a family of open causal video autoencoders built for interactive world models, the systems that generate video in response to actions, like steering a car or moving a robot arm. Many of those models borrow autoencoders from text-to-video generators, which pack many frames into one latent. That means every action waits for the chunk.

Per Kyutai's paper, world models built on CADRE run at 20 frames per second on a single H100 and respond within 230 milliseconds, roughly the point past which people notice their own movement lagging. That's the difference between a demo and something that feels controllable.

We think this is careful, useful research. The experimental design is the quiet star: seven autoencoders that differ only in the compression bottleneck, each paired with the same 1.2B world model across four environments. That isolates one variable, which is rare in a field full of apples-to-oranges comparisons.

The core insight is easy to say on air. If an autoencoder bundles eight frames at 20 fps, the bundling alone costs 400 milliseconds before anything is generated. CADRE keeps chunks short and uses a diffusion decoder distilled to one or two steps. The best-case result is striking: in driving, motion-quality FVD of 162.6 against 599.7 for the Wan 2.2 autoencoder. On latency, CADRE-2×32×32 reacts in 185 ms, versus 317 for Wan 2.2 and 605 for Cosmos; only frame-by-frame or lower-quality options come in faster.

To Kyutai's credit, the losses are on the page. In robot manipulation, Wan 2.2 follows actions far better, 0.85 against 0.45 for the most compressed CADRE, and Cosmos edges the best one-step CADRE on motion quality. Wan 2.2 also reconstructs held-out video much more faithfully. Kyutai argues reconstruction doesn't predict rollout quality, and the indoor results back that, but that's their read on their own numbers. Everything here is self-reported.

The limits matter for anyone extrapolating. Each world model covers one environment at 288×512. Whether these findings carry over to generalist world models is explicitly open. And the weights are CC BY-NC 4.0, so commercial use is off the table even though the code is MIT.

Where does this lead? Low-lag world models are a training ground for robots and self-driving systems, and potentially a new kind of game. We're for that. The risk is that convincing, responsive simulations get trusted as stand-ins for reality. A model that drives well in a generated street isn't proof it drives well on a real one, and a robot score of 0.45 on action following is a reminder of how far that can drift.

CADRE's encoder uses a frozen DINOv3 vision model plus a trained 3D convolution; its decoder is a 24-block causal diffusion transformer that streams chunk by chunk. The autoencoders were trained on about 76,000 hours from 23 public video sources spanning web, robot, driving and game footage. Kyutai released seven autoencoders, four multi-resolution versions, four world models and a keyboard-driven live demo.