title: EAGLE3 with its own draft on a mixture-of-experts model on the GB10
source: manuals/sglang-101/sections/4-4-speculative-decoding.md
lines: 112-117
lang: text
note: Divide the baseline time per token by the EAGLE-3 time per token on each row.
---
| Concurrency | Baseline median TPOT | EAGLE-3 median TPOT | Baseline output tok/s | EAGLE-3 output tok/s |
|---|---|---|---|---|
| 1 | 32.52 ms | 20.30 ms | 29.14 | 40.93 |
| 8 | 102.21 ms | 59.44 ms | 74.11 | 97.06 |
| 32 | 201.91 ms | 106.97 ms | 150.66 | 228.78 |
| 64 | 254.06 ms | 138.41 ms | 235.38 | 276.32 |
