Kyutai introduces cadre autoencoders to reduce world model latency below 230 ms

Kyutai has unveiled CADRE, a family of video autoencoders designed to enhance the responsiveness of interactive world models by reducing action lag. The system achieves a latency of under 230 ms, with a frame rate of 20 fps on a single H100 GPU. CADRE’s architecture, which includes a causal diffusion transformer decoder and a 3D convolutional encoder, is optimized for real-time performance. The autoencoders are trained on 76,000 hours of public video data, spanning web, robot, driving, and game footage. Kyutai’s models outperform existing solutions like Wan 2.2, LTX-2, and Cosmos in several benchmarks, though they face challenges in robot manipulation and indoor navigation. The code is MIT licensed, while weights are available on Hugging Face under a CC BY-NC 4.0 license. A live demo allows users to interact with the models via keyboard.