Expert Pruning Efficiency on MoE Models
0.22Calibration Time (h)AIMER
Evaluation results
| Method | Links | ||||
|---|---|---|---|---|---|
| AIMERModel=OLMoE-7B, Pruning Ratio=50%, Hardware=two NVIDIA L40S 48GB GPUs2026.03 | 0.22 | 13 | 6.89 | 12.89 | |
| AIMERModel=ERNIE-21B, Pruning Ratio=50%, Hardware=two NVIDIA L40S 48GB GPUs2026.03 | 0.51 | 40.85 | 21.67 | 40.66 | |
| REAPModel=OLMoE-7B, Pruning Ratio=50%, Hardware=two NVIDIA L40S 48GB GPUs2026.03 | 0.75 | 15.51 | 6.89 | 12.89 | |
| AIMERModel=Qwen3-30B, Pruning Ratio=50%, Hardware=two NVIDIA L40S 48GB GPUs2026.03 | 1.27 | 57.01 | 29.93 | 56.92 | |
| REAPModel=ERNIE-21B, Pruning Ratio=50%, Hardware=two NVIDIA L40S 48GB GPUs2026.03 | 1.37 | 44.72 | 21.67 | 40.66 | |
| REAPModel=Qwen3-30B, Pruning Ratio=50%, Hardware=two NVIDIA L40S 48GB GPUs2026.03 | 2.96 | 63.07 | 29.93 | 56.92 |