1-dev model. By incorporating a custom kernel optimized for 1.58-bit operations, the researchers achieved a 7.7× reduction in storage and a 5.1× reduction in inference memory usage, making deployment ...