Kyutai CADRE

CADRE (CAusal Diffusion REpresentation autoencoders) is a family of open video autoencoders built for interactive video world models, released by the open-science lab Kyutai on October 22, 2026.

Kyutai released the paper (arXiv 2610.30129, "Rethinking Video Autoencoders for Interactive World Models"), the code, the model weights and a live demo.

Why CADRE

Video world models predict the next frames from past frames and an agent's actions, and they generate inside a video autoencoder's latent space. Many borrow autoencoders from text-to-video generators, which were built for offline clip generation (no agent waiting on an action) and can't easily be retrained at a different compression. CADRE was built to study how, where and how much a world model should compress video.

Architecture

• Encoder: frozen DINOv3 ViT-B/16 features (averaged across layers)
• Bottleneck: one small trained strided convolution that sets the temporal/spatial compression
• Decoder: a space-time transformer diffusion (flow-matching) decoder, self-distilled to 1–2 steps and causal in time
• ~475M trained parameters + 86M frozen DINOv3 parameters

Key findings

• World models trained on CADRE latents run in real time at 20 fps on a single H100 and react to actions in under 230 ms, across robot and human manipulation, driving and indoor navigation
• CADRE forms most of the latency–quality frontier
• At strong compression a diffusion decoder is sharper than perceptual regression; manipulation (still camera) favors temporal compression while navigation (moving camera) favors spatial compression

How to use it

• Code: GitHub kyutai-labs/cadre-autoencoders-and-world-models
• Weights: Hugging Face kyutai/cadre-* collection (7 paper autoencoders, 4 multi-resolution variants, world models for 4 environments, action probes)
• Recommended model: kyutai/cadre-2x16x16-multires for reconstruction and new world models; the released world models pair with kyutai/cadre-2x16x16

from model_autoencoder import CADRE
autoencoder = CADRE.from_pretrained("kyutai/cadre-2x16x16")
video_hat = autoencoder.decode(autoencoder.encode(autoencoder.preprocess(video)), decode_steps=1)

Who it's for

• Researchers working on video world models, robot policy evaluation and real-time simulation/game generation
• Developers who need a low-latency video latent space

Details: Kyutai blog at kyutai.org/blog/2026-10-22-cadre/.