Inference Efficiency on MoE LLMs DSV2-16B, QW3-30B, QW3-80B-I
12.46Decode Speed (tokens/sec)BITSMOE
Evaluation Results
| Method | Links | |||||||
|---|---|---|---|---|---|---|---|---|
| BITSMOEModel=DSV2-16B2026.05 | 12.46 | 0.64 | — | — | — | 5.08 | 5.44 | |
| FP16Model=DSV2-16B2026.05 | 10.39 | 0.47 | — | 29.51 | 0.69 | 27.65 | — | |
| GPTQModel=DSV2-16B2026.05 | 7.43 | 1.27 | — | — | — | — | — | |
| BITSMOEModel=QW3-30B2026.05 | 5.71 | 1.51 | — | — | — | 8.58 | 6.29 | |
| BITSMOEModel=QW3-80B-I2026.05 | 5.01 | 1.06 | — | — | — | 21.98 | 6.56 | |
| GPTQModel=QW3-30B2026.05 | 3.25 | 2.94 | — | — | — | — | — | |
| FP16Model=QW3-30B2026.05 | 3.07 | 2.35 | — | 56.95 | 1.69 | 54 | — | |
| GPTQModel=QW3-80B-I2026.05 | 2.59 | 7.32 | — | — | — | — | — | |
| FP16Model=QW3-80B-I2026.05 | 1.65 | 8.35 | — | 148.69 | 0.61 | 144.28 | — |