Quantization Robustness Evaluation on (Wikitext, C4, HellaSwag, MMLU, Arc-C, MATH500, GSM8K) Aggregate
-0.29Accuracy Loss Delta (%)HiF8
Evaluation Results
| Method | Links | |
|---|---|---|
| HiF8Backbone=Qwen3-8B, WA Precision=W8A8, QKV Quantization=Q16KV82026.02 | -0.29 | |
| MXFP8Backbone=Qwen3-8B, WA Precision=W8A8, QKV Quantization=QKV82026.02 | -0.15 | |
| HiF8Backbone=openPangu-7B, WA Precision=W8A8, QKV Quantization=-2026.02 | -0.15 | |
| NVFP4Backbone=Qwen3-8B, WA Precision=W4A4, QKV Quantization=-2026.02 | 2.59 | |
| HiF4Backbone=Qwen3-8B, WA Precision=W4A4, QKV Quantization=Q16KV42026.02 | 3.15 |