Model Quantization ·
FP32 → INT8 → INT4
un
seel
.com · Scale · Zero-point · 4–8× smaller
Bits/weight
32
Size vs FP32
1×
State
—
FP32 weight
Snapped to grid
Outlier / clipped
Quantized model
▶ Play
←
→
🔇 Unmute
Reset
Un
seel
.com · Model Quantization