Audio Understanding on MMAU Pro
100Average ScoreVanilla
Evaluation results
| Method | Links | |||||||
|---|---|---|---|---|---|---|---|---|
| VanillaBackbone=Qwen2.5-Omni-3B, Token Pruning Level=Vanilla, Token Pruning Ratio=0%2026.04 | 100 | 59 | 54.1 | 50.6 | 65.5 | 46 | — | |
| VanillaBackbone=Qwen2.5-Omni-7B, Token Pruning Level=Vanilla, Token Pruning Ratio=0%2026.04 | 100 | 60.6 | 63.6 | 57.5 | 64.8 | 29.7 | — | |
| HeadRouterBackbone=Qwen2.5-Omni-3B, Token Pruning Level=Light, Token Pruning Ratio=30%2026.04 | 99.6 | 57.9 | 54.6 | 51.2 | 65.2 | 45 | -0.4 | |
| FrameBackbone=Qwen2.5-Omni-3B, Token Pruning Level=Light, Token Pruning Ratio=30%2026.04 | 99.1 | 57.7 | 54.6 | 50.6 | 64.9 | 44.5 | 0.9 | |
| DARTBackbone=Qwen2.5-Omni-7B, Token Pruning Level=Light, Token Pruning Ratio=30%2026.04 | 98.7 | 58.6 | 63.9 | 58.1 | 65.4 | 28 | -1.3 | |
| HeadRouterBackbone=Qwen2.5-Omni-7B, Token Pruning Level=Light, Token Pruning Ratio=30%2026.04 | 98.7 | 59 | 61.8 | 58.1 | 65.2 | 29 | -1.3 | |
| FrameBackbone=Qwen2.5-Omni-3B, Token Pruning Level=Medium, Token Pruning Ratio=60%2026.04 | 98.4 | 54.4 | 51.9 | 53.1 | 63.5 | 47.5 | -1.6 | |
| HeadRouterBackbone=Qwen2.5-Omni-3B, Token Pruning Level=Medium, Token Pruning Ratio=60%2026.04 | 98.4 | 56.7 | 53 | 52.1 | 63.6 | 45 | -1.6 | |
| FrameBackbone=Qwen2.5-Omni-7B, Token Pruning Level=Light, Token Pruning Ratio=30%2026.04 | 98.4 | 58.1 | 61.8 | 59.2 | 65.2 | 26 | 1.6 | |
| FastVBackbone=Qwen2.5-Omni-7B, Token Pruning Level=Light, Token Pruning Ratio=30%2026.04 | 97.2 | 58 | 64.1 | 56.5 | 64.3 | 27.5 | 2.8 | |
| FrameBackbone=Qwen2.5-Omni-7B, Token Pruning Level=Medium, Token Pruning Ratio=60%2026.04 | 96.9 | 56.5 | 59 | 59 | 65 | 23 | -3.1 | |
| DARTBackbone=Qwen2.5-Omni-7B, Token Pruning Level=Medium, Token Pruning Ratio=60%2026.04 | 96.7 | 52.1 | 57.4 | 59.9 | 65.7 | 27.5 | 3.3 | |
| HeadRouterBackbone=Qwen2.5-Omni-7B, Token Pruning Level=Medium, Token Pruning Ratio=60%2026.04 | 96.6 | 54.6 | 59.6 | 59.8 | 65 | 23 | 3.4 | |
| DARTBackbone=Qwen2.5-Omni-3B, Token Pruning Level=Light, Token Pruning Ratio=30%2026.04 | 96 | 55.6 | 50.3 | 49.5 | 64.4 | 44 | 4 | |
| FastVBackbone=Qwen2.5-Omni-7B, Token Pruning Level=Medium, Token Pruning Ratio=60%2026.04 | 95.9 | 55.4 | 57.6 | 57.3 | 64 | 28 | 4.1 | |
| DARTBackbone=Qwen2.5-Omni-3B, Token Pruning Level=Medium, Token Pruning Ratio=60%2026.04 | 91.8 | 48.2 | 48.6 | 50.2 | 63.4 | 42 | 8.2 | |
| HeadRouterBackbone=Qwen2.5-Omni-3B, Token Pruning Level=Extreme, Token Pruning Ratio=90%2026.04 | 91.6 | 46.5 | 44.3 | 55 | 60.5 | 45.5 | -8.4 | |
| FastVBackbone=Qwen2.5-Omni-3B, Token Pruning Level=Light, Token Pruning Ratio=30%2026.04 | 91.5 | 52.5 | 47 | 42.4 | 58.6 | 51 | 8.5 | |
| FrameBackbone=Qwen2.5-Omni-7B, Token Pruning Level=Extreme, Token Pruning Ratio=90%2026.04 | 90.2 | 42.4 | 49.2 | 58.8 | 62.5 | 26.5 | -9.8 | |
| HeadRouterBackbone=Qwen2.5-Omni-7B, Token Pruning Level=Extreme, Token Pruning Ratio=90%2026.04 | 90.2 | 41.7 | 50.3 | 58.8 | 62.9 | 25.5 | -9.8 | |
| FastVBackbone=Qwen2.5-Omni-3B, Token Pruning Level=Medium, Token Pruning Ratio=60%2026.04 | 90 | 49.8 | 46.4 | 41.3 | 58.2 | 51.5 | 10 | |
| DARTBackbone=Qwen2.5-Omni-7B, Token Pruning Level=Extreme, Token Pruning Ratio=90%2026.04 | 89.8 | 41.4 | 50.3 | 58.5 | 61.5 | 26 | 10.2 | |
| FrameBackbone=Qwen2.5-Omni-3B, Token Pruning Level=Extreme, Token Pruning Ratio=90%2026.04 | 89.6 | 42.3 | 41 | 54.3 | 61.1 | 48 | 10.4 | |
| FastVBackbone=Qwen2.5-Omni-7B, Token Pruning Level=Extreme, Token Pruning Ratio=90%2026.04 | 88.8 | 44.7 | 53.7 | 57 | 58.9 | 23 | 11.2 | |
| DARTBackbone=Qwen2.5-Omni-3B, Token Pruning Level=Extreme, Token Pruning Ratio=90%2026.04 | 88.5 | 41 | 41.5 | 53.5 | 61.7 | 45.5 | 11.5 | |
| FastVBackbone=Qwen2.5-Omni-3B, Token Pruning Level=Extreme, Token Pruning Ratio=90%2026.04 | 84.5 | 41.7 | 44.3 | 39.5 | 55.3 | 52 | 15.5 | |
| Gemini-3.1-ProSize=–, Model Category=Closed-source models2026.06 | 73.47 | — | — | — | — | — | — | |
| Gemini-3-ProSize=–, Model Category=Closed-source models2026.06 | 68.28 | — | — | — | — | — | — | |
| MOSS-Audio-8B-ThinkingSize=8B, Model Category=Open-source models2026.06 | 64.92 | — | — | — | — | — | — | |
| Qwen3-Omni-30B-A3B-InstructSize=30B, Model Category=Open-source models2026.06 | 61.22 | — | — | — | — | — | — | |
| Step-Audio-R1.1Size=33B, Model Category=Open-source models2026.06 | 60.8 | — | — | — | — | — | — | |
| MOSS-Audio-4B-ThinkingSize=4B, Model Category=Open-source models2026.06 | 60.75 | — | — | — | — | — | — | |
| Step-Audio-R1Size=33B, Model Category=Open-source models2026.06 | 59.68 | — | — | — | — | — | — | |
| MOSS-Audio-4B-InstructSize=4B, Model Category=Open-source models2026.06 | 58.16 | — | — | — | — | — | — | |
| MOSS-Audio-8B-InstructSize=8B, Model Category=Open-source models2026.06 | 57.48 | — | — | — | — | — | — | |
| Kimi-AudioSize=7B, Model Category=Open-source models2026.06 | 56.58 | — | — | — | — | — | — | |
| Audio Flamingo NextSize=8B, Model Category=Open-source models2026.06 | 56.34 | — | — | — | — | — | — | |
| MiMo-Audio-7BSize=7B, Model Category=Open-source models2026.06 | 53.35 | — | — | — | — | — | — | |
| GPT4o-AudioSize=–, Model Category=Closed-source models2026.06 | 52.3 | — | — | — | — | — | — | |
| Qwen2.5-OmniSize=7B, Model Category=Open-source models2026.06 | 52.2 | — | — | — | — | — | — | |
| Audio Flamingo 3Size=7B, Model Category=Open-source models2026.06 | 51.7 | — | — | — | — | — | — | |
| MiniCPM-o-4.5Size=9B, Model Category=Open-source models2026.06 | 39.65 | — | — | — | — | — | — |