Video Understanding on VideoMME
100Overall ScoreBaseline
Evaluation Results
| Method | Links | ||||||||||||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| BaselineBackbone=Qwen2.5-VL-7B-Instruct, Token Retention=25%2026.02 | 100 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | 61.33 | — | |
| IDPrunerBackbone=Qwen2.5-VL-7B-Instruct, Token Retention=25%2026.02 | 87.13 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | 59.48 | — | |
| DivPruneBackbone=Qwen2.5-VL-7B-Instruct, Token Retention=25%2026.02 | 87.06 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | 58 | — | |
| GPT-5Size=-, # Frames=-2026.01 | 86.7 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| SCOPEBackbone=Qwen2.5-VL-7B-Instruct, Token Retention=25%2026.02 | 86.4 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | 60 | — | |
| VisionZipBackbone=Qwen2.5-VL-7B-Instruct, Token Retention=25%2026.02 | 85.98 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | 59.67 | — | |
| DARTBackbone=Qwen2.5-VL-7B-Instruct, Token Retention=25%2026.02 | 85.19 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | 59.52 | — | |
| FastVBackbone=Qwen2.5-VL-7B-Instruct, Token Retention=25%2026.02 | 84.56 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | 59.44 | — | |
| HiPruneBackbone=Qwen2.5-VL-7B-Instruct, Token Retention=25%2026.02 | 84.41 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | 59.93 | — | |
| Gemini-2.5-ProSize=-, # Frames=-2026.01 | 84.3 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| VisPrunerBackbone=Qwen2.5-VL-7B-Instruct, Token Retention=25%2026.02 | 83.45 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | 59.44 | — | |
| VisionSelectorBackbone=Qwen2.5-VL-7B-Instruct, Token Retention=25%2026.02 | 81.81 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | 59.19 | — | |
| Gemini-1.5-ProParams=-, Frames=0.5 fps, Subtitle Setting=w/ S2024.11 | 81.3 | — | — | — | — | — | — | — | — | 77.4 | 84.5 | 81 | — | — | — | — | — | — | |
| Seed1.5-VLAccess=Proprietary2026.01 | 77.9 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| LLaVA-Video + Video-RAGParams=72B, Frames=64, Subtitle Setting=Ours (Video-RAG)2024.11 | 77.4 | — | — | — | — | — | — | — | — | 73.1 | 82.8 | 76.3 | 1.5 | — | — | — | — | — | |
| GPT-4oParams=-, Frames=384, Subtitle Setting=w/ S2024.11 | 77.2 | — | — | — | — | — | — | — | — | 72.1 | 82.8 | 76.6 | — | — | — | — | — | — | |
| Doubao-Seed-1.6Access=Proprietary2026.01 | 76.5 | — | — | — | — | — | — | — | — | 70.1 | 84.1 | 75.3 | — | — | — | — | — | — | |
| LLaVA-Video + TV-RAG#Text=2.1K, LLM Params=72B, Frames=322025.12 | 76.5 | — | — | — | — | — | — | — | — | 73.8 | 82.3 | 73.1 | 9.4 | — | — | — | — | — | |
| LLaVA-VideoParams=72B, Frames=64, Subtitle Setting=w/ S2024.11 | 75.9 | — | — | — | — | — | — | — | — | 72.2 | 81.8 | 73.8 | — | — | — | — | — | — | |
| Gemini-1.5-ProParams=-, Frames=0.5 fps, Subtitle Setting=w/o S2024.11 | 75 | — | — | — | — | — | — | — | — | 67.4 | 81.7 | 74.3 | — | — | — | — | — | — | |
| Gemini-1.5-Pro2025.03 | 75 | — | — | — | — | — | — | — | — | 67.4 | — | — | — | — | — | — | — | — | |
| Gemini-1.5-ProSize=-, # Frames=-2026.01 | 75 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| Gemini-1.5-ProFrames=0.5 fps2025.12 | 75 | — | — | — | — | — | — | — | — | 67.4 | 81.7 | 74.3 | — | — | — | — | — | — | |
| Qwen2-VL + TV-RAG#Text=2.1K, LLM Params=72B, Frames=322025.12 | 74.9 | — | — | — | — | — | — | — | — | 72.1 | 76.3 | 70.5 | 10 | — | — | — | — | — | |
| Qwen2.5-VL+AdaRETAKELLM Size=72B2025.03 | 73.5 | — | — | — | — | — | — | — | — | 65 | — | — | — | — | — | — | — | — | |
| Qwen2-VL + Video-RAGParams=72B, Frames=32, Subtitle Setting=Ours (Video-RAG)2024.11 | 72.9 | — | — | — | — | — | — | — | — | 71 | 77.4 | 70.2 | 1 | — | — | — | — | — | |
| Qwen2.5-VLLLM Size=72B2025.03 | 72.6 | — | — | — | — | — | — | — | — | 63.9 | — | — | — | — | — | — | — | — | |
| InternVL2.5LLM Size=72B2025.03 | 72.1 | — | — | — | — | — | — | — | — | 62.6 | — | — | — | — | — | — | — | — | |
| GPT-4oParams=-, Frames=384, Subtitle Setting=w/o S2024.11 | 71.9 | — | — | — | — | — | — | — | — | 65.3 | 80 | 70.3 | — | — | — | — | — | — | |
| Qwen2-VLParams=72B, Frames=32, Subtitle Setting=w/ S2024.11 | 71.9 | — | — | — | — | — | — | — | — | 69.2 | 76.7 | 69.9 | — | — | — | — | — | — | |
| GPT-4o2025.03 | 71.9 | — | — | — | — | — | — | — | — | 65.3 | — | — | — | — | — | — | — | — | |
| GPT-4oSize=-, # Frames=-2026.01 | 71.9 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| GPT-4oAccess=Proprietary2026.01 | 71.9 | — | — | — | — | — | — | — | — | 65.3 | 80 | 70.3 | — | — | — | — | — | — | |
| GPT-4oFrames=3842025.12 | 71.9 | — | — | — | — | — | — | — | — | 65.3 | 80 | 70.3 | — | — | — | — | — | — | |
| FullBackbone=Qwen3-VL-8B2026.02 | 71.22 | — | — | — | — | — | — | — | — | 63.11 | 79.89 | 70.67 | — | — | — | — | — | — | |
| PrismBackbone=Qwen3-VL-8B2026.02 | 71.22 | — | — | — | — | — | — | — | — | 64 | 79 | 70.67 | — | — | — | — | — | — | |
| Qwen2-VLLLM Size=72B2025.03 | 71.2 | — | — | — | — | — | — | — | — | 62.2 | — | — | — | — | — | — | — | — | |
| XAttentionBackbone=Qwen3-VL-8B2026.02 | 70.81 | — | — | — | — | — | — | — | — | 63.44 | 79.22 | 69.78 | — | — | — | — | — | — | |
| MInferenceBackbone=Qwen3-VL-8B2026.02 | 70.63 | — | — | — | — | — | — | — | — | 62.44 | 79.44 | 70 | — | — | — | — | — | — | |
| LLaVA-VideoLLM Size=72B2025.03 | 70.6 | — | — | — | — | — | — | — | — | 61.5 | — | — | — | — | — | — | — | — | |
| Qwen3-Omni-30B-A3BSize=30B2026.02 | 70.5 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| FlexPrefillBackbone=Qwen3-VL-8B2026.02 | 70.34 | — | — | — | — | — | — | — | — | 62.67 | 77.67 | 70.67 | — | — | — | — | — | — | |
| LLaVA-VideoParams=72B, Frames=64, Subtitle Setting=w/o S2024.11 | 70.3 | — | — | — | — | — | — | — | — | 62.1 | 80.7 | 68.7 | — | — | — | — | — | — | |
| LiVi-LLM-7BAccess=Open-Source2026.01 | 68.5 | — | — | — | — | — | — | — | — | 56.6 | 79.6 | 69.4 | — | — | — | — | — | — | |
| OmniVinciSize=7B2026.02 | 68.2 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| Qwen2.5-VL+AdaRETAKELLM Size=7B2025.03 | 67.7 | — | — | — | — | — | — | — | — | 58.3 | — | — | — | — | — | — | — | — | |
| Qwen2.5-Omni-7BAccess=Open-Source2026.01 | 67.7 | — | — | — | — | — | — | — | — | 57.6 | 78.1 | 67.4 | — | — | — | — | — | — | |
| AriaLLM Size=8x3.5B2025.03 | 67.6 | — | — | — | — | — | — | — | — | 58.8 | — | — | — | — | — | — | — | — | |
| Oryx-1.5LLM Size=32B2025.03 | 67.3 | — | — | — | — | — | — | — | — | 59.3 | — | — | — | — | — | — | — | — | |
| LLaVA-VideoLLM Params=72B, Frames=322025.12 | 67.1 | — | — | — | — | — | — | — | — | 59.6 | 78 | 63.7 | — | — | — | — | — | — | |
| LLaVA-OneVisionLLM Size=72B2025.03 | 66.3 | — | — | — | — | — | — | — | — | 60 | — | — | — | — | — | — | — | — | |
| InternVL3-8BAccess=Open-Source2026.01 | 66.3 | — | — | — | — | — | — | — | — | 54.1 | 77.5 | 67.3 | — | — | — | — | — | — | |
| VideoLLaMA3LLM Size=7B2025.03 | 66.2 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| TimeLens-7B2025.12 | 65.7 | — | — | — | — | — | — | — | — | 54.1 | 66.4 | 76.7 | — | — | — | — | — | — | |
| TPOLLM Size=7B2025.03 | 65.6 | — | — | — | — | — | — | — | — | 55.4 | — | — | — | — | — | — | — | — | |
| Qwen2.5-VL-7BAccess=Open-Source2026.01 | 65.6 | — | — | — | — | — | — | — | — | 54.1 | 75.9 | 66.8 | — | — | — | — | — | — | |
| Qwen2.5-VLLLM Size=7B2025.03 | 65.4 | — | — | — | — | — | — | — | — | 55.6 | — | — | — | — | — | — | — | — | |
| SCORERetention Ratio R=25%, Max Frames=1282026.03 | 65.3 | — | — | — | — | — | — | — | — | 55.7 | 74.8 | 65.4 | — | 102.8 | — | — | — | 58.9 | |
| Qwen2.5-VLSize=7B, # Frames=-2026.01 | 65.1 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| Qwen2.5-VL-7Bvariant=Baseline2025.12 | 65.1 | — | — | — | — | — | — | — | — | 55.1 | 64.3 | 75.2 | — | — | — | — | — | — | |
| SCORERetention Ratio R=40%, Max Frames=1282026.03 | 65 | — | — | — | — | — | — | — | — | 55.1 | 75 | 65 | — | 102.4 | — | — | — | 58.7 | |
| Qwen2-VLParams=72B, Frames=32, Subtitle Setting=w/o S2024.11 | 64.9 | — | — | — | — | — | — | — | — | 56.3 | 75 | 63.3 | — | — | — | — | — | — | |
| Qwen2-VLLLM Params=72B, Frames=322025.12 | 64.9 | — | — | — | — | — | — | — | — | 56.3 | 75 | 63.3 | — | — | — | — | — | — | |
| DyCokeRetention Ratio R=40%, Max Frames=1282026.03 | 64.9 | — | — | — | — | — | — | — | — | 55 | 74.3 | 65.3 | — | 101.7 | — | — | — | 58.2 | |
| VidCom^2Retention Ratio R=40%, Max Frames=1282026.03 | 64.7 | — | — | — | — | — | — | — | — | 54.6 | 74.8 | 64.7 | — | 102.1 | — | — | — | 58.5 | |
| ByteVideoLLMLLM Size=14B2025.03 | 64.6 | — | — | — | — | — | — | — | — | 56.4 | — | — | — | — | — | — | — | — | |
| VidCom^2Retention Ratio R=25%, Max Frames=1282026.03 | 64.6 | — | — | — | — | — | — | — | — | 55.3 | 73.3 | 65.2 | — | 101.6 | — | — | — | 58.2 | |
| Qwen2.5-OmniSize=7B2026.02 | 64.3 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| LLaVA-Video-7BPrefilling FLOPs (T)=80.2, FLOPs Ratio=100%, Before LLM Retained Ratio=100%2026.03 | 64.3 | — | — | — | — | — | — | — | — | — | — | — | — | 100 | — | — | — | 60.2 | |
| NVILALLM Size=8B2025.03 | 64.2 | — | — | — | — | — | — | — | — | 54.8 | — | — | — | — | — | — | — | — | |
| Qwen2-VL+AdaRETAKELLM Size=7B2025.03 | 64.2 | — | — | — | — | — | — | — | — | 56.4 | — | — | — | — | — | — | — | — | |
| InternVL2.5-8BAccess=Open-Source2026.01 | 64.2 | — | — | — | — | — | — | — | — | 55.8 | 75.3 | 61.5 | — | — | — | — | — | — | |
| LLaVA-Video+AdaRETAKELLM Size=7B2025.03 | 64 | — | — | — | — | — | — | — | — | 53.9 | — | — | — | — | — | — | — | — | |
| OmniFysicsSize=3B2026.02 | 63.8 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| VanillaRetention Ratio R=100%, Max Frames=1282026.03 | 63.8 | — | — | — | — | — | — | — | — | 52.8 | 74.4 | 64.2 | — | 100 | — | — | — | 57.3 | |
| FastVIDRetention Ratio R=40%, Max Frames=1282026.03 | 63.7 | — | — | — | — | — | — | — | — | 54.4 | 73.9 | 62.8 | — | 100.3 | — | — | — | 57.5 | |
| SCORERetention Ratio R=10%, Max Frames=1282026.03 | 63.6 | — | — | — | — | — | — | — | — | 53.9 | 73.6 | 63.1 | — | 99.5 | — | — | — | 57 | |
| LLaVA-VideoLLM Size=7B2025.03 | 63.3 | — | — | — | — | — | — | — | — | 52.4 | — | — | — | — | — | — | — | — | |
| Qwen2-VLLLM Size=7B2025.03 | 63.3 | — | — | — | — | — | — | — | — | 53.8 | — | — | — | — | — | — | — | — | |
| FastVIDRetention Ratio R=25%, Max Frames=1282026.03 | 63.3 | — | — | — | — | — | — | — | — | 53.9 | 74.3 | 61.7 | — | 99.8 | — | — | — | 57.1 | |
| Video-RTSSize=7B, # Frames=51.22026.01 | 63 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| Long-LLaVA + Video-RAGParams=7B, Frames=64, Subtitle Setting=Ours (Video-RAG)2024.11 | 62.6 | — | — | — | — | — | — | — | — | 60.1 | 67.1 | 60.4 | 4.8 | — | — | — | — | — | |
| Video-KTRSize=7B, # Frames=642026.01 | 62.5 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| AOTPrefilling FLOPs (T)=9.3, FLOPs Ratio=18.9%, Before LLM Retained Ratio=25%2026.03 | 62.4 | — | — | — | — | — | — | — | — | — | — | — | — | 96.8 | — | — | — | 58.3 | |
| Long-LLaVA + TV-RAG#Text=1.9K, LLM Params=7B, Frames=322025.12 | 62.1 | — | — | — | — | — | — | — | — | 59.8 | 66.4 | 60.2 | 10.1 | — | — | — | — | — | |
| LongVA + Video-RAGParams=7B, Frames=128, Subtitle Setting=Ours (Video-RAG)2024.11 | 62 | — | — | — | — | — | — | — | — | 59.4 | 66.1 | 60.4 | 6 | — | — | — | — | — | |
| Qwen2.5-OmniSize=3B2026.02 | 62 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| FastVIDRetention Ratio R=10%, Max Frames=1282026.03 | 62 | — | — | — | — | — | — | — | — | 51.4 | 72.9 | 61.7 | — | 95.9 | — | — | — | 54.9 | |
| PDropPrefilling FLOPs (T)=19.5, FLOPs Ratio=24.3%, Before LLM Retained Ratio=100%2026.03 | 61.9 | — | — | — | — | — | — | — | — | — | — | — | — | 94.2 | — | — | — | 56.7 | |
| AOTPrefilling FLOPs (T)=9.3, FLOPs Ratio=11.6%, Before LLM Retained Ratio=15%2026.03 | 61.9 | — | — | — | — | — | — | — | — | — | — | — | — | 95.1 | — | — | — | 57.3 | |
| VidCom^2Retention Ratio R=10%, Max Frames=1282026.03 | 61.8 | — | — | — | — | — | — | — | — | 52.6 | 70.4 | 62.4 | — | 95.3 | — | — | — | 54.6 | |
| LongVA + TV-RAG#Text=1.8K, LLM Params=7B, Frames=322025.12 | 61.7 | — | — | — | — | — | — | — | — | 58.7 | 66.2 | 62.1 | 10.3 | — | — | — | — | — | |
| HoliTomRetention Ratio R=40%, Max Frames=1282026.03 | 61.6 | — | — | — | — | — | — | — | — | 51.8 | 73.1 | 59.9 | — | 99.3 | — | — | — | 56.9 | |
| DyCokeRetention Ratio R=25%, Max Frames=1282026.03 | 61 | — | — | — | — | — | — | — | — | 51.7 | 71.2 | 60 | — | 94.8 | — | — | — | 54.3 | |
| VisionZipPrefilling FLOPs (T)=9.3, FLOPs Ratio=18.9%, Before LLM Retained Ratio=25%2026.03 | 60.7 | — | — | — | — | — | — | — | — | — | — | — | — | 94.2 | — | — | — | 56.7 | |
| VisionZipPrefilling FLOPs (T)=9.3, FLOPs Ratio=11.6%, Before LLM Retained Ratio=15%2026.03 | 60.7 | — | — | — | — | — | — | — | — | — | — | — | — | 94.2 | — | — | — | 56.7 | |
| Our Final ModelTraining Stages=PT, RL, Decoding Mode=concise reason2025.12 | 60.6 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| Qwen2.5-VL-3BBase Model=Qwen2.5-VL-3B, Equivalent Retention Ratio=100%2026.02 | 60.4 | — | — | — | — | — | — | — | — | 49.6 | 71.4 | 60.1 | — | — | — | — | — | — | |
| Video-KTRSize=7B, # Frames=322026.01 | 60.3 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| Qwen3-VL-8BTraining Stages=PT, Decoding Mode=direct answer2025.12 | 60.3 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — |