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