Video Multimodal Understanding on VideoMME Overall
116.9Prefilling Latency (s)Random
Evaluation results
| Method | Links | ||||||
|---|---|---|---|---|---|---|---|
| Randommax input frames=32, R (token/budget ratio)=25%2026.03 | 116.9 | 147.3 | 1,271.9 | 19,547.5 | 2.12 | 61.5 | |
| V-CASTmax input frames=32, R (token/budget ratio)=25%2026.03 | 118.8 | 149.9 | 1,245.1 | 19,547.5 | 2.17 | 63.5 | |
| VidCom2max input frames=32, R (token/budget ratio)=25%2026.03 | 119 | 150.2 | 1,271 | 19,547.5 | 2.12 | 62.4 | |
| FastVIDmax input frames=32, R (token/budget ratio)=25%2026.03 | 120.6 | 153.6 | 1,317.6 | 19,547.5 | 2.05 | 60.5 | |
| HoliTommax input frames=32, R (token/budget ratio)=25%2026.03 | 134.3 | 160.8 | 1,261.8 | 19,974.7 | 2.14 | 59.7 | |
| FlashVIDmax input frames=32, R (token/budget ratio)=25%2026.03 | 146.6 | 175 | 1,308.7 | 41,363.6 | 2.06 | 62.3 | |
| VisionZipmax input frames=32, R (token/budget ratio)=25%2026.03 | 155.6 | 190.5 | 1,276.8 | 19,974.7 | 2.11 | 60.1 | |
| Qwen3-VL-8B-Instructmax input frames=32, R (token/budget ratio)=25%2026.03 | 243.6 | 280.9 | 1,440.4 | 22,478 | 1.87 | 64.5 |