Kyutai releases CADRE video autoencoders to reduce action lag in interactive world modelsMachine translation
Automatically verified and published · Generated and evidence-checked automatically; not reviewed by a human.
Kyutai has released CADRE video autoencoders and model weights for developers building action-conditioned world models. Its paper reports that CADRE world models run at 20 fps on one H100 and respond within 230 ms. The figures are self-reported by Kyutai, and the weights are licensed for non-commercial use.
This source permits summary display only.
Read at the original source