Kyutai’s CADRE combines frozen DINOv3 features with a causal video decoder to reduce response latency in world models for tasks such as manipulation and driving. Its repository reports 20 fps and 162 ms local-input latency for one user on an H100, falling to about 14 fps per user with two users. These latency measurements exclude network transit and browser rendering, and long rollouts can drift. Code is MIT-licensed, but weights use CC-BY-NC 4.0, so commercial-use rights cannot be assumed identical.