List your company

Quantization Distribution Evaluation on Wiki2 (calibration set)

0KL DivergenceBF16 (baseline)
-0.0057160.0328670.071450.110033May 12, 2026

Evaluation results

MethodLinks
BF16 (baseline)Precision=BF162026.050
PO2(Split87)Quantization=W4A42026.050.0813
PO2(NF4)Quantization=W4A42026.050.0893
Split87Quantization=W4A42026.050.0897
BOF4Quantization=W4A42026.050.0969
NF4Quantization=W4A42026.050.0988
IF4Quantization=W4A42026.050.0992
SFP4Quantization=W4A42026.050.1019
4over6Quantization=W4A42026.050.1023
NVFP4Quantization=W4A42026.050.1143
NVINT4Quantization=W4A42026.050.1429