1 related articles
Kyutai REKI 2.4B is an open-source vision-language-action model that uses 32 register tokens to implement short-term memory inside a 200M action expert, achieving 21.8ms inference on a single H100 and 90.4% on LIBERO-Plus. This article analyzes its Flow Matching architecture, register caching mechanism, benchmark performance, and LeRobot deployment path.
