Real-world Question Answering on RWQA
76.99RWQA AccuracyQwen3-VL-235B
Evaluation Results
| Method | Links | |
|---|---|---|
| Qwen3-VL-235BToken Retention Ratio=100%2026.05 | 76.99 | |
| F3AToken Retention Ratio=60%2026.05 | 76.86 | |
| DivPruneToken Retention Ratio=60%2026.05 | 75.94 | |
| Qwen3-VL-32BToken Ratio=100%2026.05 | 75.82 | |
| CDPrunerToken Retention Ratio=60%2026.05 | 75.82 | |
| VisionZipToken Retention Ratio=60%2026.05 | 75.66 | |
| F3AToken Retention Ratio=40%2026.05 | 75.21 | |
| FastVToken Retention Ratio=60%2026.05 | 74.98 | |
| FLAMODEL=4B, Precision=FP2026.06 | 74.77 | |
| F3AToken Ratio=60%2026.05 | 74.64 | |
| VisionZipToken Retention Ratio=40%2026.05 | 74.21 | |
| CDPrunerToken Retention Ratio=40%2026.05 | 74.11 | |
| DivPruneToken Retention Ratio=40%2026.05 | 74.02 | |
| F3AToken Ratio=40%2026.05 | 73.33 | |
| FLAMODEL=4B, Precision=W4A162026.06 | 73.33 | |
| F3AToken Retention Ratio=20%2026.05 | 73.2 | |
| FastVToken Retention Ratio=40%2026.05 | 72.98 | |
| Multiplication-Only Matrix Inversion ApproximationMODEL=4B, Precision=W4A162026.06 | 72.81 | |
| SGT-BAGELParams=7B+7B‡2026.05 | 72.42 | |
| DivPruneToken Ratio=60%2026.05 | 71.63 | |
| DivPruneToken Retention Ratio=20%2026.05 | 71.42 | |
| BAGELParams=7B+7B‡2026.05 | 71.26 | |
| CDPrunerToken Ratio=60%2026.05 | 71.11 | |
| Molmo 7B# Vis tok.=12002025.04 | 71.1 | |
| VisionZipToken Retention Ratio=20%2026.05 | 71.03 | |
| CDPrunerToken Retention Ratio=20%2026.05 | 70.85 | |
| VisionZipToken Ratio=60%2026.05 | 70.07 | |
| FastVToken Ratio=60%2026.05 | 69.02 | |
| CDPrunerToken Ratio=40%2026.05 | 68.76 | |
| Qwen2.5VL 7B# Vis tok.=14002025.04 | 68.4 | |
| DivPruneToken Ratio=40%2026.05 | 68.1 | |
| VisionZipToken Ratio=20%2026.05 | 67.71 | |
| UniWorld-v1Params=7B+12B‡2026.05 | 67.58 | |
| FLARE-X 8B# Vis tok.=14002025.04 | 66.9 | |
| FastVToken Retention Ratio=20%2026.05 | 66.72 | |
| VisionZipToken Ratio=40%2026.05 | 66.54 | |
| Eagle 8B# Vis tok.=10242025.04 | 66.5 | |
| UniMRGParams=3.6B2026.05 | 66.01 | |
| LLaVA-OneVision 7B# Vis tok.=24002025.04 | 65.5 | |
| Qwen2.5VL 3B# Vis tok.=14002025.04 | 65.4 | |
| FLARE-L 8B# Vis tok.=6302025.04 | 65.4 | |
| OpenUniParams=2B2026.05 | 65.23 | |
| SGT-Gen2Params=3B+4B‡2026.05 | 65.1 | |
| OmniGen2Params=3B+4B‡2026.05 | 64.41 | |
| DeepSeek-VL2-Tiny# Vis tok.=15002025.04 | 64.2 | |
| Cambrain 8B# Vis tok.=5762025.04 | 64.2 | |
| Florence-VL 8B# Vis tok.=5762025.04 | 64.2 | |
| UniLIPParams=2B2026.05 | 64.18 | |
| FLARE-X 3B# Vis tok.=14002025.04 | 63.7 | |
| MiniCPM-Llama3-V-2.5 8B# Vis tok.=4002025.04 | 63.5 | |
| FastVToken Ratio=40%2026.05 | 62.75 | |
| DivPruneToken Ratio=20%2026.05 | 62.35 | |
| FLAMODEL=0.8B, Precision=FP2026.06 | 62.35 | |
| FLARE-L 3B# Vis tok.=6302025.04 | 62.1 | |
| CDPrunerToken Ratio=20%2026.05 | 61.96 | |
| F3AToken Ratio=20%2026.05 | 61.83 | |
| Qwen2-VL-2B#Params=2B, TFLOPs=0.42025.06 | 60.7 | |
| LaTtE-Flow#Params=2B, TFLOPs=0.42025.06 | 60.7 | |
| Florence-VL 3B# Vis tok.=5762025.04 | 60.4 | |
| Multiplication-Only Matrix Inversion ApproximationMODEL=0.8B, Precision=W4A162026.06 | 60.39 | |
| MetaMorph#Params=8B, TFLOPs=1.12025.06 | 58.3 | |
| Emu3Params=8B2026.05 | 57.4 | |
| EMU3#Params=8B, TFLOPs=4.12025.06 | 57.4 | |
| FLAMODEL=0.8B, Precision=W4A162026.06 | 56.08 | |
| MiniCPM-V-2.0 3B# Vis tok.=4002025.04 | 55.8 | |
| FastVToken Ratio=20%2026.05 | 53.99 | |
| Phi 3.5-Vision# Vis tok.=21002025.04 | 53.5 | |
| HarmonParams=1.5B2026.05 | 48 | |
| VILA-U#Params=7B, TFLOPs=3.62025.06 | 46.6 | |
| ReCA-HarmonParams=1.5B2026.05 | 43.53 | |
| Janus-ProParams=7B2026.05 | 41.83 | |
| Chameleon#Params=34B, TFLOPs=17.42025.06 | 39.2 | |
| ChameleonParams=7B2026.05 | 39 | |
| Chameleon#Params=7B, TFLOPs=3.62025.06 | 39 | |
| Show-o512Params=1.3B2026.05 | 38.17 |