Quantization formats

A quantization label names a mixture, not a uniform precision — so the effective bits per weight of a real file always exceeds the nominal rate its name implies. These are the measured averages across every file in the index.

FormatFiles measuredEffective bpw
UD-TQ1_0502.067
I1-IQ1_S9572.068
IQ1_S1922.075
UD-IQ1_S1572.160
IQ1_M2552.170
I1-IQ1_M9562.209
UD-IQ1_M1782.311
IQ2_XXS3422.393
I1-IQ2_XXS9572.445
UD-IQ2_XXS1992.591
IQ2_XS4472.614
I1-IQ2_XS9572.647
IQ2_S4522.712
I1-IQ2_S9572.762
TQ2_0192.808
UD-IQ2_M2062.937
I1-IQ2_M9572.950
IQ2_M7452.988
I1-Q2_K_S8293.030
Q2_K_S863.184
I1-Q2_K10373.223
UD-IQ3_S323.250
TQ1_0143.256
I1-IQ3_XXS9593.278
UD-IQ3_XXS2133.298
Q2_K23313.337
Q2_K_L7643.379
IQ3_XXS5423.427
Q2_K_M83.451
I1-IQ3_XS9603.518
IQ3_XS7883.562
I1-Q3_K_S10363.680
I1-IQ3_S10363.698
Q3_K_S22413.744
I1-IQ3_M10343.796
IQ3_S2353.867
IQ3_M9063.871
UD-IQ4_XS363.989
UD-IQ4_NL294.012
I1-Q3_K_M10354.029
Q3_K_M24614.127
I1-Q3_K_L10344.321
Q3_K_L21344.368
UD-Q3_K_M374.412
I1-IQ4_XS10344.413
Q3_K844.493
IQ4_XS19684.556
I1-Q4_010334.625
I1-Q4_K_S10344.672
I1-IQ4_NL5284.683
UD-Q3_K_S104.737
MXFP4134.761
IQ4_NL8404.779
Q4_K_S23604.795
Q4_013764.913
I1-Q4_K_M10334.925
I1-Q4_18975.051
Q4_K_M33125.134
Q4_K_L5655.203
UD-Q4_K_S285.285
Q4_18785.371
I1-Q5_K_S10165.536
Q5_K_S23195.656
I1-Q5_K_M10185.691
Q5_K_M27765.919
UD-Q4_K_M325.947
Q5_K_L5275.964
Q4_K1646.044
Q5_05066.111
UD-Q5_K_S286.255
Q5_K1086.374
Q6_K_M106.415
I1-Q6_K9906.541
UD-Q6_K286.740
UD-Q2_K56.751
Q6_K29996.808
Q6_K_L5266.811
NVFP4306.849
Q5_12686.942
UD-Q5_K_M317.032
Q8_034518.635
Q2_078.699
BF1694515.545
F16116115.704
F3215831.257