Rest assured, all the big players (openai, google, deepseek etc) have run countless experiments with 4,3,2,1.58,1 bits, and various sparse factors and shapes. This barrel has been scraped to the bottom