Kyutai releases REKI, a 2.4B vision-language-action model with cached memory

Kyutai, with Inria and École des Ponts, has released REKI, a 2.4B vision-language-action model for robot manipulation, pretrained only on public data.

Its memory stores 32 cached register tokens across the last three frames, adding 0.1 ms to a 21.8 ms action chunk on one H100. REKI scores 90.4% on LIBERO-Plus, the top score in Kyutai's comparison, and its weights are on Hugging Face under CC BY-NC-SA 4.0.