List your company

Language Modeling on Llama-2 7B base (PPL, KL)

4.7527PerplexityBF16 (ref)
4.7159444.9640475.212155.460253May 25, 2026

Evaluation results

MethodLinks
BF16 (ref)bpw=16.002026.054.7527——
SmoothQuant W8A8bpw=8.132026.054.75760.10.0009
QAM-W-5.5bpw=5.512026.054.76640.290.0043
QTIP 4-bit (relaxml)bpw=4.002026.054.81551.320.01
QAM-W-4bpw=4.002026.054.85512.150.0292
QTIP 2-bit (relaxml)bpw=2.002026.055.584117.490.1195
AQLM 2-bit 1×16 (ISTA-DASLab)bpw=2.032026.055.671619.330.1337