Meet Kyutai REKI: The 2.4B VLA Model Bringing Efficient Memory to Robot Manipulation
MarkTechPost 2 days ago 39
Kyutai released REKI, a 2.4B vision-language-action model for robot manipulation that uses cached memory of recent frames. The model runs at 21.8 ms per compiled action chunk on an NVIDIA H100. REKI’s built-in 3-frame memory adds 0.1 ms and reduces pose oscillations on RoboTwin from 18.6% to 4.1% while reaching 90.4% on LIBERO-Plus.