Speculative Decoding Acceleration on OmniDocBench (8192 tokens v1.5)
4.98AAL (Overall)Qwen3-VL-8B
Evaluation Results
| Method | Links | ||||||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Qwen3-VL-8BParameters=9B, Method=Hierarchical Speculative Decoding2026.02 | 4.98 | 2.61 | 2.62 | 1.63 | 2.56 | 2.29 | 2.02 | 2.18 | 2.59 | 4.62 | 1.86 | 1.91 | |
| Qwen3-VL-2BParameters=2B, Method=Hierarchical Speculative Decoding2026.02 | 4.33 | 2.18 | 2.18 | 1.35 | 2.13 | 1.82 | 1.87 | 2.04 | 2.61 | 4.03 | 1.69 | 1.75 |