Video Understanding on Video-MME (Accuracy)
84.8AccuracyGemini-2.5 Pro
Evaluation results
| Method | Links | |
|---|---|---|
| Gemini-2.5 ProAgent=disabled2026.05 | 84.8 | |
| GPT-4oAgent=disabled2026.05 | 77.2 | |
| Gemini-1.5-ProLLM Size=-, Frames=-2026.07 | 75 | |
| Qwen2.5-VL-72BLLM Size=72B, Frames=1fps2026.07 | 73.3 | |
| GPT-4oLLM Size=-, Frames=256 / 0.5fps2026.07 | 71.9 | |
| QCALLM Size=30B, Frames=64, Backbone=Qwen3-VL-30B-A3B2026.07 | 71.4 | |
| CapRL-Video-178KBackbone=Molmo2-8B2026.06 | 71.1 | |
| Tarsier2-7B-Video-178KBackbone=Molmo2-8B2026.06 | 70.9 | |
| InternVL3.5-38BLLM Size=38B, Frames=642026.07 | 70.9 | |
| GPT-4o-Video-178KBackbone=Molmo2-8B2026.06 | 70.5 | |
| ShareGPT4Video-8B-Video-178KBackbone=Molmo2-8B2026.06 | 70.3 | |
| CapRL-Video-178KBackbone=Molmo2-4B2026.06 | 70.2 | |
| LLaVA-Video-72BLLM Size=72B, Frames=642026.07 | 70 | |
| Qwen3-VL-30B-A3BLLM Size=30B, Frames=642026.07 | 69.9 | |
| Tarsier2-7B-Video-178KBackbone=Molmo2-4B2026.06 | 69.7 | |
| QCALLM Size=8B, Frames=64, Backbone=Qwen3-VL-8B2026.07 | 69.5 | |
| GPT-4o-Video-178KBackbone=Molmo2-4B2026.06 | 68.7 | |
| ShareGPT4Video-8B-Video-178KBackbone=Molmo2-4B2026.06 | 68.4 | |
| VideoSeeker-8BAgent=enabled2026.05 | 68.1 | |
| Kimi-VL-16B-A3BLLM Size=16B, Frames=642026.07 | 67.8 | |
| Qwen3-VL-8BLLM Size=8B, Frames=642026.07 | 67.6 | |
| Qwen3-VL-8BAgent=disabled2026.05 | 67.4 | |
| Qwen3-VL-8B + ADPOPost-training Strategy=ADPO2026.05 | 66.7 | |
| InternVL3-8B + HY-HimmelCtx (k)=16.22026.05 | 66.3 | |
| LongVT-7BAgent=disabled2026.05 | 66.1 | |
| VideoSeeker-4BAgent=enabled2026.05 | 66.1 | |
| Qwen3-VL-8B + DAPOPost-training Strategy=DAPO2026.05 | 65.9 | |
| Qwen3-VL-8B + GRPOPost-training Strategy=GRPO2026.05 | 65.4 | |
| Qwen3-VL-4BAgent=disabled2026.05 | 65.3 | |
| AKSFrames=642025.10 | 65.3 | |
| Qwen3-VL-8B + HY-HimmelCtx (k)=16.22026.05 | 64.9 | |
| Qwen3-VL-8BPost-training Strategy=Base Model2026.05 | 64.7 | |
| LLaVA-Video-7BFrames per video=64, Retain Tokens=All 64 x 169 (100%)2026.07 | 64.4 | |
| InternVL3-8B baselineCtx (k)=44.72026.05 | 64.3 | |
| NVILALLM Size=7B, Frames=2562026.07 | 64 | |
| Qwen2.5-VL-7B + HY-HimmelCtx (k)=16.22026.05 | 63.5 | |
| Qwen3-VL-8B baselineCtx (k)=44.72026.05 | 62.7 | |
| EADPFrames per video=64, Retain Tokens=64 x 64 (↓ 62.1%)2026.07 | 62 | |
| Video-SFTBackbone=Qwen2.5-VL-7B, Frames=64F2026.03 | 61.93 | |
| Hybrid-Frame StrategyBackbone=Qwen2.5-VL-7B, Frames=32F2026.03 | 61.52 | |
| Qwen3-VL-4BAgent=enabled2026.05 | 61.5 | |
| Video-SFTBackbone=Qwen2.5-VL-7B, Frames=32F2026.03 | 61.3 | |
| ApolloLLM Size=7B, Frames=2fps2026.07 | 61.3 | |
| Qwen2.5-VL-7B baselineCtx (k)=44.72026.05 | 61.2 | |
| DivPruneFrames per video=64, Retain Tokens=64 x 64 (↓ 62.1%)2026.07 | 61.2 | |
| Video-R1-7BAgent=disabled2026.05 | 61 | |
| Video-Thinker-7BAgent=disabled2026.05 | 61 | |
| Hybrid-Frame StrategyBackbone=Qwen2.5-VL-7B, Frames=64F2026.03 | 60.93 | |
| VideoRFT-7BAgent=disabled2026.05 | 60.9 | |
| CDPrunerFrames per video=64, Retain Tokens=64 x 64 (↓ 62.1%)2026.07 | 60.9 | |
| EADPFrames per video=64, Retain Tokens=64 x 32 (↓ 81.1%)2026.07 | 60.4 | |
| Video-SFTBackbone=LLaVA-1.5-7B, Frames=32F2026.03 | 59.9 | |
| LLaVA-OV-7B + HY-HimmelCtx (k)=16.22026.05 | 59.6 | |
| mPLUG-Owl3LLM Size=8B, Frames=1282026.07 | 59.3 | |
| DivPruneFrames per video=64, Retain Tokens=64 x 32 (↓ 81.1%)2026.07 | 59.2 | |
| HiPruneFrames per video=64, Retain Tokens=64 x 64 (↓ 62.1%)2026.07 | 59 | |
| LLaVA-Video-7B + FrameOracleFrames=64→82025.10 | 58.9 | |
| Video-SFTBackbone=Qwen2.5-VL-7B, Frames=16F2026.03 | 58.74 | |
| LLaVA-Video-7B + BOLTFrames=1fps→82025.10 | 58.6 | |
| Qwen3-VL-8BAgent=enabled2026.05 | 58.3 | |
| CDPrunerFrames per video=64, Retain Tokens=64 x 32 (↓ 81.1%)2026.07 | 58.3 | |
| Hybrid-Frame StrategyBackbone=Qwen2.5-VL-7B, Frames=16F2026.03 | 58.15 | |
| LLaVA-OV-7B baselineCtx (k)=44.72026.05 | 58 | |
| LLaVA-Video-7B + KFCFrames=1fps→82025.10 | 57.6 | |
| LLaVA-OneVision-7B + Frame-VoyagerFrames=128→82025.10 | 57.5 | |
| LLaVA-OneVision-7B + FrameOracleFrames=64→82025.10 | 57.5 | |
| Qwen2.5-VL-7B + K-framesFrames=256→82025.10 | 57.4 | |
| Qwen2.5-VL-7B + FrameOracleFrames=128→82025.10 | 57.4 | |
| FP16 (Full Precision)W/A Bits=16/162026.06 | 57.33 | |
| Qwen2.5-VL-7B + ViaRLFrames=128→82025.10 | 57.3 | |
| EADPFrames per video=64, Retain Tokens=64 x 16 (↓ 90.5%)2026.07 | 57.3 | |
| DivPruneFrames per video=64, Retain Tokens=64 x 16 (↓ 90.5%)2026.07 | 57.2 | |
| CDPrunerFrames per video=64, Retain Tokens=64 x 16 (↓ 90.5%)2026.07 | 56.3 | |
| LLaVA-OneVision-7B + BOLTFrames=1fps→82025.10 | 56.1 | |
| HiPruneFrames per video=64, Retain Tokens=64 x 32 (↓ 81.1%)2026.07 | 56 | |
| LLaVA-Video-7BFrames=82025.10 | 55.9 | |
| LLaVA-OneVision-7B + KFCFrames=1fps→82025.10 | 55.4 | |
| Hybrid-Frame StrategyBackbone=Qwen2.5-VL-7B, Frames=8F2026.03 | 54.96 | |
| QLoRAW/A Bits=4/162026.06 | 54.78 | |
| Video-SFTBackbone=Qwen2.5-VL-7B, Frames=8F2026.03 | 54.41 | |
| Qwen2.5-Omni (Ours)W/A Bits=4/42026.06 | 54.33 | |
| Qwen2.5-VL-7BFrames=82025.10 | 54.1 | |
| LLaVA-OneVision-7BFrames=82025.10 | 53.8 | |
| HiPruneFrames per video=64, Retain Tokens=64 x 16 (↓ 90.5%)2026.07 | 53.8 | |
| Hybrid-Frame StrategyBackbone=LLaVA-1.5-7B, Note=Evaluated with a fixed 8-frame inference2026.03 | 52.41 | |
| Video-SFTBackbone=LLaVA-1.5-7B, Frames=16F2026.03 | 51.39 | |
| PoMQ-ViTW/A Bits=4/42026.06 | 49.22 | |
| QVLMW/A Bits=4/42026.06 | 48.41 | |
| Naive W4A4W/A Bits=4/42026.06 | 47.52 | |
| Video-SFTBackbone=LLaVA-1.5-7B, Frames=8F2026.03 | 38.59 |