Weight Quantization Fidelity on Llama 3B 3.2
2.11MSEBlock-scaled FP6
Evaluation Results
| Method | Links | |
|---|---|---|
| Block-scaled FP6Weight Format=E2M3sUE4M4, Bits per weight (bpw)=6.5, Scaling Granularity=per-block2026.05 | 2.11 | |
| per-layer-POT FP8Weight Format=E4M3ˆ0sUE8M0, Bits per weight (bpw)=8.0, Scaling Granularity=per-layer POT2026.05 | 2.72 |