TurboQuant can reduce vector index size by 10x at 100M Row Scale (github.com)

[dead]

32 bits vs 4 bits it looks like