End-to-End Inference Performance on LongBench single-node
0.79TTFT (Mean)FoE
Evaluation Results
| Method | Links | |||||||||
|---|---|---|---|---|---|---|---|---|---|---|
| FoEScale=1.5, Architecture=FoE2026.05 | 0.79 | 0.744 | 1.627 | 0.334 | 0.333 | 0.365 | 86 | 85.6 | 94.1 | |
| FoEScale=0.75, Architecture=FoE2026.05 | 1.041 | 0.979 | 2.116 | 0.458 | 0.458 | 0.502 | 117.9 | 117.8 | 129 | |
| FoEScale=0.5, Architecture=FoE2026.05 | 1.394 | 1.213 | 3.528 | 0.566 | 0.596 | 0.639 | 145.7 | 153.8 | 164.3 | |
| MoEScale=1.5, Architecture=MoE2026.05 | 2.264 | 2.118 | 5.128 | 0.371 | 0.36 | 0.558 | 96.9 | 93.8 | 146.7 | |
| MoEScale=0.75, Architecture=MoE2026.05 | 3.233 | 3.147 | 5.856 | 0.841 | 0.869 | 1.033 | 217.7 | 225.2 | 266.6 | |
| MoEScale=0.5, Architecture=MoE2026.05 | 3.71 | 3.623 | 6.437 | 0.78 | 0.783 | 0.979 | 202.7 | 203.1 | 253.3 |