Quantization, in the terms this site uses

Quantisation reduces the bits used per weight. Fewer bits means less memory, and because decoding is bandwidth-bound, proportionally faster generation.

The site records quant (the label a backend uses, such as Q4_K_M) and quant_bpw (the effective bits per weight, 4.90 for Q4_K_M). The label alone is not enough for arithmetic: Q4_K_M is not exactly 4 bits, because k-quant formats mix precisions across tensors.

Two rows at different quantisations are not comparable, which is why the quantisation is part of every record’s identity. A 4-bit row being faster than an 8-bit row on the same hardware is arithmetic, not a finding.

Memory saved is not free. Lower bit widths cost output quality, and the loss is not linear. This site does not publish quality measurements, so it deliberately makes no claim about which quantisation is “good enough” — it only reports what each one costs and what speed it produced.