aiqbench analysis

Kyutai's REKI is a 2.4B vision-language-action model for robot manipulation, notable for training on public data only and for a low-cost memory of recent observations, with its weights released openly. Its claim of outperforming models roughly twice its size on three out-of-distribution benchmarks is promising but depends on benchmark details and independent replication. For AI readers, the practical questions are how the memory mechanism scales over longer tasks and how it performs on real hardware outside the reported evaluations.