Video Understanding on FAVOR-Bench
50.28ASQwen2.5-VL-72B*
Evaluation results
| Method | Links | ||||||||
|---|---|---|---|---|---|---|---|---|---|
| Qwen2.5-VL-72B*Frames=1fps, Model Size Series=Large Size Series2026.06 | 50.28 | 46.98 | 48.13 | 51.78 | 40.28 | 51.56 | 48.14 | 48.17 | |
| Qwen2.5VL-7B + MotionEnhancerFrames=1fps, Model Size Series=Medium Size Series2026.06 | 49.34 | 50.62 | 45.37 | 53.2 | 30.42 | 53.12 | 46.88 | 47.01 | |
| LLaVA-Video-72B-Qwen2Frames=64, Model Size Series=Large Size Series2026.06 | 48.35 | 47.5 | 45.25 | 51.7 | 33.02 | 53.12 | 46.08 | 46.49 | |
| InternVL3-8B + MotionEnhancerFrames=1fps, Model Size Series=Medium Size Series2026.06 | 47.17 | 57.11 | 46.57 | 56.35 | 36.74 | 51.56 | 48.94 | 49.25 | |
| VideoLLaMA3-2B + FAVOR-Train + MOSSTraining=Fine-tuned on FAVOR-Train, Module=MOSS2026.04 | 46.8 | 55 | 45.7 | 52.2 | 28.8 | 53.3 | 46.6 | — | |
| Qwen2.5VL-7B + MotionSight*Frames=1fps, Model Size Series=Medium Size Series2026.06 | 46.23 | 51.59 | 45.01 | 50.04 | 29.95 | 53.12 | 45.47 | 45.99 | |
| InternVL3-8B*Frames=1fps, Model Size Series=Medium Size Series2026.06 | 45.39 | 48.54 | 47.59 | 51.45 | 33.58 | 51.56 | 45.82 | 46.35 | |
| Qwen2.5VL-3B + MotionEnhancerFrames=1fps, Model Size Series=Small Size Series2026.06 | 45.01 | 51.59 | 44.4 | 48.96 | 28.37 | 45.31 | 44.53 | 43.94 | |
| VideoLLaMA3-2B + FAVOR-TrainTraining=Fine-tuned on FAVOR-Train2026.04 | 44.8 | 54.5 | 45.3 | 51.3 | 27.8 | 54.7 | 45.5 | — | |
| VideoLLaMA3-2BTraining=Zero-shot / Pre-trained2026.04 | 43.8 | 44.4 | 42.8 | 48.1 | 27.3 | 45.3 | 42.2 | — | |
| VideoChat-Flash-Qwen2-7BFrames=1fps, Model Size Series=Medium Size Series2026.06 | 41.9 | 48.41 | 42.84 | 50.95 | 35.07 | 50 | 43.82 | 44.86 | |
| Qwen2.5VL-7B*Frames=1fps, Model Size Series=Medium Size Series2026.06 | 41.64 | 47.83 | 44.89 | 47.55 | 28.28 | 45.31 | 42.61 | 42.58 | |
| VideoLLaMA3-7BFrames=1fps, Model Size Series=Medium Size Series2026.06 | 40.2 | 44.13 | 42.42 | 48.3 | 31.53 | 42.19 | 41.46 | 41.46 | |
| Full CacheModel=Qwen2.5-VL-7B-Instruct2025.11 | 39.63 | 43.61 | 43.2 | 43.57 | 33.67 | 40.62 | 40.91 | — | |
| InternVL3-2B + MotionEnhancerFrames=1fps, Model Size Series=Small Size Series2026.06 | 38.53 | 54.57 | 42.6 | 51.78 | 33.02 | 51.56 | 43.71 | 45.35 | |
| Qwen2.5VL-3B*Frames=1fps, Model Size Series=Small Size Series2026.06 | 38.45 | 38.16 | 39.35 | 43.4 | 23.72 | 45.31 | 37.43 | 38.07 | |
| InternVL2.5-78BFrames=8, Model Size Series=Large Size Series2026.06 | 38.38 | 40.62 | 39.05 | 43.65 | 29.4 | 39.06 | 38.54 | 38.36 | |
| InternVL3-2B*Frames=1fps, Model Size Series=Small Size Series2026.06 | 37.66 | 43.28 | 40.49 | 44.98 | 29.21 | 39.06 | 39.27 | 39.11 | |
| LLaVA-Video-7B-Qwen2Frames=64, Model Size Series=Medium Size Series2026.06 | 36.14 | 41.27 | 41.28 | 44.48 | 29.58 | 46.88 | 38.6 | 39.94 | |
| H2OModel=Qwen2.5-VL-7B-Instruct, KV Cache retention ratio ρ=0.12025.11 | 32.16 | 38.22 | 39.05 | 40.66 | 31.35 | 23.44 | 35.78 | — | |
| LLaVA-NeXT-Video-34BFrames=8, Model Size Series=Large Size Series2026.06 | 31.7 | 31.99 | 32.31 | 22.99 | 29.58 | 46.88 | 30.44 | 32.58 | |
| StreamingLLMModel=Qwen2.5-VL-7B-Instruct, KV Cache retention ratio ρ=0.12025.11 | 30.94 | 39.13 | 39.05 | 40.5 | 31.35 | 28.12 | 35.57 | — | |
| LOOK-MModel=Qwen2.5-VL-7B-Instruct, KV Cache retention ratio ρ=0.12025.11 | 30.38 | 40.23 | 39.53 | 43.4 | 32.56 | 25 | 36.25 | — | |
| FlashCacheModel=Qwen2.5-VL-7B-Instruct, KV Cache retention ratio ρ=0.12025.11 | 30.38 | 40.1 | 41.16 | 40.91 | 34.14 | 35.94 | 36.49 | — | |
| VideoLLaMA3-2BFrames=1fps, Model Size Series=Small Size Series2026.06 | 28.97 | 36.6 | 34.9 | 38.01 | 28.56 | 40.62 | 32.98 | 34.61 | |
| SnapKVModel=Qwen2.5-VL-7B-Instruct, KV Cache retention ratio ρ=0.12025.11 | 28.9 | 39.71 | 40.91 | 39.83 | 33.49 | 39.06 | 35.67 | — | |
| MEDAModel=Qwen2.5-VL-7B-Instruct, KV Cache retention ratio ρ=0.12025.11 | 26.17 | 36.28 | 37.61 | 39.75 | 31.91 | 21.88 | 33.11 | — | |
| Video-LLaVA-7BFrames=8, Model Size Series=Medium Size Series2026.06 | 24.91 | 21.54 | 25.45 | 30.54 | 26.23 | 21.88 | 25.37 | 25.09 | |
| SnapKVModel=Qwen2.5-VL-7B-Instruct, KV Cache retention ratio ρ=0.052025.11 | 24.19 | 32.25 | 34.66 | 34.69 | 29.02 | 15.62 | 29.95 | — | |
| H2OModel=Qwen2.5-VL-7B-Instruct, KV Cache retention ratio ρ=0.052025.11 | 21.35 | 28.68 | 32.19 | 31.45 | 28.56 | 12.5 | 27.3 | — | |
| FlashCacheModel=Qwen2.5-VL-7B-Instruct, KV Cache retention ratio ρ=0.052025.11 | 21.27 | 34.78 | 37.55 | 34.69 | 33.21 | 26.56 | 30.71 | — | |
| LLaVA-NeXT-Video-7BFrames=8, Model Size Series=Medium Size Series2026.06 | 21.27 | 22.45 | 26.05 | 26.72 | 23.07 | 14.06 | 23.45 | 22.27 | |
| StreamingLLMModel=Qwen2.5-VL-7B-Instruct, KV Cache retention ratio ρ=0.052025.11 | 20.52 | 29.4 | 32.55 | 31.29 | 28 | 12.5 | 27.14 | — | |
| LOOK-MModel=Qwen2.5-VL-7B-Instruct, KV Cache retention ratio ρ=0.052025.11 | 20.29 | 30.11 | 32.49 | 33.61 | 31.81 | 12.5 | 28.03 | — | |
| InternVL2.5-2BFrames=8, Model Size Series=Small Size Series2026.06 | 18.7 | 28.23 | 23.71 | 27.47 | 19.16 | 23.44 | 22.9 | 23.45 | |
| MEDAModel=Qwen2.5-VL-7B-Instruct, KV Cache retention ratio ρ=0.052025.11 | 15.17 | 24.72 | 25.99 | 29.13 | 27.53 | 12.5 | 22.83 | — | |
| Tarsier-7BFrames=8, Model Size Series=Medium Size Series2026.06 | 12.55 | 21.16 | 17.87 | 17.93 | 22.23 | 31.25 | 17.46 | 20.5 |