End-to-End Latency Evaluation on Dataset A (N=5,504)
10,169Median Latency (p50)always-ds4
Evaluation Results
| Method | Links | ||||
|---|---|---|---|---|---|
| always-ds4Method Type=Single-model baseline, Latency Component=LLM generation latency only2026.06 | 10,169 | 135,296 | 468,067 | 41,768 | |
| always-qwenMethod Type=Single-model baseline, Latency Component=LLM generation latency only2026.06 | 21,441 | 232,232 | 266,774 | 52,677 | |
| Brick (MoM)Method Type=Mixture-of-Models, Profile=max profile, Latency Component=Router decision plus LLM generation2026.06 | 22,802 | 278,621 | 820,288 | 78,178 | |
| always-kimiMethod Type=Single-model baseline, Latency Component=LLM generation latency only2026.06 | 51,246 | 390,276 | 738,106 | 106,926 |