Speculative Decoding on 16-Dataset Aggregate Suite Prompts (test)
104.5Throughput (tokens/s)Frequency Hybrid
Evaluation Results
| Method | Links | |||||
|---|---|---|---|---|---|---|
| Frequency HybridTarget Model=Qwen3-4B2026.05 | 104.5 | 223.9 | 3.47 | 8.2 | 91.8 | |
| Hybrid VerifiedTarget Model=Qwen3-4B2026.05 | 101.1 | 228.6 | 3.37 | 8.8 | 91.2 | |
| EAGLE3Target Model=Qwen3-4B2026.05 | 76.7 | 309.7 | 2.23 | 0 | 100 | |
| SuffixTarget Model=Qwen3-4B2026.05 | 54.6 | 346.1 | 1.89 | 97.6 | 0 | |
| GreedyTarget Model=Qwen3-4B2026.05 | 38.3 | 1,000 | — | — | — |