Kyutai unveils CADRE, saying it cuts world-model action latency to under 230 ms
French AI research lab Kyutai has introduced CADRE, a family of open causal diffusion representation autoencoders built for interactive world models. The release is aimed at reducing action latency, a key constraint for systems that need to react quickly to user input or changing environments. According to Kyutai, its world model runs at 20 fps on a single NVIDIA H100 GPU and delivers response times below 230 milliseconds. The model family includes seven autoencoders, each with about 476 million trainable parameters, and uses a frozen 86 million-parameter DINOv3 ViT-B/16 model. Kyutai said the full world model totals 1.2 billion parameters. In benchmark comparisons against MIRA, Wan 2.2, Cosmos, and LTX-2, CADRE outperformed rivals in 42 of 48 tests. In a driving-scene FVD result, CADRE-2×16×16 with single-step decoding posted 162.6, compared with 599.7 for Wan 2.2. The system was trained on a mixed dataset spanning 23 public video sources and roughly 76,000 hours of data, including web video, robotics, driving, and gameplay footage.








