Multimodal Understanding on MME (Score)
2,393ScoreInternVL3
Evaluation Results
| Method | Links | |
|---|---|---|
| InternVL3Params=8B2026.04 | 2,393 | |
| VisionZipTokens Retained Percentage=33.3%2025.03 | 2,387 | |
| TwigVLM++Tokens Retained Percentage=33.3%2025.03 | 2,373 | |
| Qwen3-VLParams=8B2026.04 | 2,364 | |
| InternVL3.5Params=8B2026.04 | 2,357 | |
| Qwen2.5-VL-7B2025.12 | 2,347 | |
| Q2.5VL-7BTokens Retained Percentage=100%2025.03 | 2,347 | |
| TwigVLM++Tokens Retained Percentage=22.2%2025.03 | 2,346 | |
| InternVL2.5-8B2025.12 | 2,344 | |
| Qwen2.5-OmniParams (B)=7, Time (s)=6.02025.08 | 2,340 | |
| InternVL2.5-4BComplexity=O(n)2025.12 | 2,338 | |
| TwigVLMTokens Retained Percentage=33.3%2025.03 | 2,338 | |
| Qwen2-VL-7B2025.12 | 2,326 | |
| DART + RandomBackbone=Qwen2.5-VL-7B, Token Retention Ratio=50%2025.12 | 2,318 | |
| VisionZipTokens Retained Percentage=22.2%2025.03 | 2,318 | |
| Qwen2.5-VLParams=7B2026.04 | 2,317 | |
| Qwen2.5-VL-7BBackbone=Qwen2.5-VL-7B, Token Retention Ratio=100%2025.12 | 2,313 | |
| GPT-4oTime (s)=1.22025.08 | 2,310 | |
| DivPrune+VTWBackbone=Qwen2.5-VL-7B, Token Retention Ratio=50%2025.12 | 2,308 | |
| DART+VTWBackbone=Qwen2.5-VL-7B, Token Retention Ratio=50%2025.12 | 2,302 | |
| DivPrune + RandomBackbone=Qwen2.5-VL-7B, Token Retention Ratio=50%2025.12 | 2,302 | |
| FastVTokens Retained Percentage=33.3%2025.03 | 2,299 | |
| DARTBackbone=Qwen2.5-VL-7B, Token Retention Ratio=50%2025.12 | 2,295 | |
| DivPruneBackbone=Qwen2.5-VL-7B, Token Retention Ratio=50%2025.12 | 2,291 | |
| CoT4DET-7B2025.12 | 2,289 | |
| Qwen-VL-Max2025.08 | 2,281 | |
| FastVTokens Retained Percentage=22.2%2025.03 | 2,246 | |
| TwigVLM++Tokens Retained Percentage=11.1%2025.03 | 2,240 | |
| TwigVLMTokens Retained Percentage=22.2%2025.03 | 2,238 | |
| InternVL2-8B2025.12 | 2,210 | |
| Qwen2.5VL-3BComplexity=O(n)2025.12 | 2,171 | |
| VisionZipTokens Retained Percentage=11.1%2025.03 | 2,153 | |
| InfiniteVL-4BComplexity=O(1)2025.12 | 2,126 | |
| TwigVLMTokens Retained Percentage=11.1%2025.03 | 2,020 | |
| VITA1.5Params (B)=7, Time (s)=3.72025.08 | 2,006 | |
| LLaVA-OV-7B2025.12 | 1,998 | |
| Qwen2.5-VLParams (B)=722025.08 | 1,980 | |
| PAIBackbone=Qwen-VL-Chat2026.05 | 1,923 | |
| Ours:base Qwen2.5Params (B)=14+72, Time (s)=3.22025.08 | 1,922 | |
| VDD-NoneBackbone=Qwen-VL-Chat2026.05 | 1,921 | |
| Qwen2.5-VLParams (B)=322025.08 | 1,915 | |
| BRACSBackbone=Qwen-VL-Chat2026.05 | 1,882 | |
| BRACSBackbone=LLaVA-1.5-7B2026.05 | 1,880 | |
| UniDriveVLAParams=8B2026.04 | 1,876 | |
| GreedyBackbone=LLaVA-1.5-7B2026.05 | 1,866 | |
| VanillaToken Budget (K)=576, Token Reduction (%)=100%, Model Backbone=LLaVA-1.5-7B2026.06 | 1,864 | |
| VanillaEarly Cmp.=−, Retained Tokens=576, Compression Ratio=100%2026.03 | 1,862 | |
| VCDBackbone=Qwen-VL-Chat2026.05 | 1,860 | |
| FastVTokens Retained Percentage=11.1%2025.03 | 1,859 | |
| VCDBackbone=LLaVA-1.5-7B2026.05 | 1,853 | |
| Phi-3.5-Vision-4BComplexity=O(n)2025.12 | 1,846 | |
| PriorTRToken Budget (K)=192, Token Reduction (%)=↓ 66.7%, Model Backbone=LLaVA-1.5-7B2026.06 | 1,845 | |
| SPINBackbone=Qwen-VL-Chat2026.05 | 1,824 | |
| PriorTRToken Budget (K)=128, Token Reduction (%)=↓ 77.8%, Model Backbone=LLaVA-1.5-7B2026.06 | 1,820 | |
| GreedyBackbone=Qwen-VL-Chat2026.05 | 1,808 | |
| PRUNESID-DynEarly Cmp.=✓, Retained Tokens=192, Compression Ratio=66.7%2026.03 | 1,797 | |
| VisPrunerToken Budget (K)=192, Token Reduction (%)=↓ 66.7%, Model Backbone=LLaVA-1.5-7B2026.06 | 1,796 | |
| PAIBackbone=LLaVA-1.5-7B2026.05 | 1,775 | |
| VDD-NoneBackbone=LLaVA-1.5-7B2026.05 | 1,772 | |
| MaTVLM-3BComplexity=O(n)2025.12 | 1,771 | |
| SPINBackbone=LLaVA-1.5-7B2026.05 | 1,768 | |
| PDropToken Budget (K)=192, Token Reduction (%)=↓ 66.7%, Model Backbone=LLaVA-1.5-7B2026.06 | 1,766 | |
| SmolVLM2-2BComplexity=O(n)2025.12 | 1,764 | |
| VisPrunerToken Budget (K)=128, Token Reduction (%)=↓ 77.8%, Model Backbone=LLaVA-1.5-7B2026.06 | 1,761 | |
| PRUNESID-DynEarly Cmp.=✓, Retained Tokens=128, Compression Ratio=77.8%2026.03 | 1,760 | |
| PriorTRToken Budget (K)=64, Token Reduction (%)=↓ 88.9%, Model Backbone=LLaVA-1.5-7B2026.06 | 1,745 | |
| PRUNESID-DynEarly Cmp.=✓, Retained Tokens=64, Compression Ratio=88.9%2026.03 | 1,734 | |
| TinyLLaVA-3BComplexity=O(n)2025.12 | 1,733 | |
| SparseVLMToken Budget (K)=192, Token Reduction (%)=↓ 66.7%, Model Backbone=LLaVA-1.5-7B2026.06 | 1,721 | |
| SparseVLMToken Budget (K)=128, Token Reduction (%)=↓ 77.8%, Model Backbone=LLaVA-1.5-7B2026.06 | 1,696 | |
| Bagel2026.04 | 1,687 | |
| BAGALType=Und., # Params=14B2025.03 | 1,687 | |
| VisPrunerToken Budget (K)=64, Token Reduction (%)=↓ 88.9%, Model Backbone=LLaVA-1.5-7B2026.06 | 1,687 | |
| BLIP3o 8B2026.04 | 1,683 | |
| Qwen2.5-VLParams (B)=72025.08 | 1,673 | |
| PaliGemma2-3BComplexity=O(n)2025.12 | 1,658 | |
| LatentUMBasequantized setting=false2026.04 | 1,654 | |
| PDropToken Budget (K)=128, Token Reduction (%)=↓ 77.8%, Model Backbone=LLaVA-1.5-7B2026.06 | 1,644 | |
| LatentUMBasequantized setting=true2026.04 | 1,638 | |
| LLaVA-NeXTType=Und., # Params=34B2025.03 | 1,631 | |
| PruMergeToken Budget (K)=192, Token Reduction (%)=↓ 66.7%, Model Backbone=LLaVA-1.5-7B2026.06 | 1,626 | |
| DualToken-7BType=Uni., # Params=7B, Resolution=384px2025.03 | 1,625 | |
| Show-o2Type=Uni., # Params=7B, Resolution=432px2025.03 | 1,620.5 | |
| Show-o22026.04 | 1,620 | |
| FastVToken Budget (K)=192, Token Reduction (%)=↓ 66.7%, Model Backbone=LLaVA-1.5-7B2026.06 | 1,605 | |
| DualToken-3BType=Uni., # Params=3B, Resolution=384px2025.03 | 1,588.4 | |
| ShareGPT4VType=Und., # Params=7B2025.03 | 1,567.4 | |
| Tokenflow XL2026.04 | 1,551 | |
| PruMergeToken Budget (K)=64, Token Reduction (%)=↓ 88.9%, Model Backbone=LLaVA-1.5-7B2026.06 | 1,549 | |
| VIFLLM=Vicuna-7B, Resolution=336 × 3362026.04 | 1,547.2 | |
| TokenFlowType=Uni., # Params=13B, Resolution=384px2025.03 | 1,545.9 | |
| PruMergeToken Budget (K)=128, Token Reduction (%)=↓ 77.8%, Model Backbone=LLaVA-1.5-7B2026.06 | 1,544 | |
| VILAType=Und., # Params=7B2025.03 | 1,533 | |
| BLIP3o 4B2026.04 | 1,528 | |
| LLaVA-HRLLM=Vicuna-7B, Resolution=384 × 3842026.04 | 1,522.3 | |
| IGVALLM=Vicuna-7B, Resolution=336 × 3362026.04 | 1,519.8 | |
| LLaVA-NeXTLLM=Mistral-7B, Resolution=672 × 6722026.04 | 1,519.3 | |
| LLaVA-NeXTType=Und., # Params=7B2025.03 | 1,519 | |
| LLaVA-1.5Type=Und., # Params=7B2025.03 | 1,510.7 | |
| Original*Backbone=LLaVA-1.5-7B, Visual Token #=576, TFLOPs (ratio %)=2.23 (100)2026.03 | 1,506.5 |