So i assumed it would get crushed by OPQ (which requires training)

you’re right that 32f is faster on raw query time, quantization adds extra step. main benefit on download size since gzip won’t help much, which matters most in browser contexts