Kyutai releases open REKI 2.4B vision-language-action model

France's Kyutai, with Inria and École des Ponts, released REKI, an open 2.4B-parameter vision-language-action model with 32 learned register tokens for short-term memory. It scores 90.4% on LIBERO-Plus and cuts robotic arm oscillation from 18.6% to 4.1% on RoboTwin.

• 2.4B parameters: 2.2B vision-language backbone plus 0.2B Flow Matching action expert
• 32 register tokens compress 3 seconds of observations at 2 MiB VRAM per frame
• 21.8 ms inference on a single H100, registers add just 0.1 ms
• 90.4% success on LIBERO-Plus, oscillation cut from 18.6% to 4.1%