Diagram Understanding on AI2D (test)
94.7AccuracyClaude-3.5-Sonnet
Evaluation Results
| Method | Links | |
|---|---|---|
| Claude-3.5-Sonnet2024.12 | 94.7 | |
| Gemini-1.5-Pro2024.12 | 94.4 | |
| GPT-4o2024.12 | 94.2 | |
| Molmo2-8BParameter Scale=8B2026.05 | 94.1 | |
| Molmo2-4BParameter Scale=4B2026.05 | 93.8 | |
| InternVL3.5-8BParameter Scale=8B2026.05 | 92.5 | |
| Qwen3-VL-8BParameter Scale=8B2026.05 | 92.3 | |
| PerceptionLM-3BParameter Scale=3B2026.05 | 92.2 | |
| InternVL3.5-4BParameter Scale=4B2026.05 | 92 | |
| Qwen3-VL-4BParameter Scale=4B2026.05 | 91.8 | |
| PerceptionLM-8BParameter Scale=8B2026.05 | 91.8 | |
| Zamba2-VL-7BParameter Scale=7B2026.05 | 90.6 | |
| Gemini-2.5-ProDate=2024.12, Tokenization Paradigm=Proprietary Models2026.02 | 89.5 | |
| BagelModel Scale=14B, Date=2025.05, Tokenization Paradigm=Continuous Tokenization2026.02 | 89.2 | |
| Claude-3 Sonnetopen-source=false2024.04 | 88.7 | |
| InternVL3.5-2BParameter Scale=2B2026.05 | 88.6 | |
| InternVL2-76BModel Scale=76B2024.12 | 88.4 | |
| Grok-1.5Vopen-source=false2024.04 | 88.3 | |
| Claude-3 Opusopen-source=false2024.04 | 88.1 | |
| Qwen-VL-Max-0809Params (B)=722024.10 | 88.1 | |
| Qwen2-VL-72BModel Scale=72B2024.12 | 88.1 | |
| InternVL2-Llama3-76BParams (B)=762024.10 | 87.6 | |
| Qwen2-VL-72B-ThinkingMax Output Tokens=40K2026.03 | 87.2 | |
| Qwen2-VL-72B-ThinkingMax Output Tokens=4K2026.03 | 86.9 | |
| Claude-3 Haikuopen-source=false2024.04 | 86.7 | |
| Qwen3-VL-2BParameter Scale=2B2026.05 | 86.2 | |
| ManzanoModel Scale=30B, Date=2025.09, Tokenization Paradigm=Hybrid Tokenization2026.02 | 86 | |
| Zamba2-VL-2.7BParameter Scale=2.7B2026.05 | 85.8 | |
| PerceptionLM-1BParameter Scale=1B2026.05 | 85.7 | |
| LLaVA-OV-72BModel Scale=72B2024.12 | 85.6 | |
| LLaVA-OV-72BModel Scale=72B, Single-image variant (SI)=true2024.12 | 85.1 | |
| VQ-RAEModel Scale=7B, Date=2025.11, Tokenization Paradigm=Hybrid Tokenization2026.02 | 84.8 | |
| Phi-4-reasoning-vision-15BThinking Mode=default2026.03 | 84.8 | |
| GPT-4o-202408062024.10 | 84.7 | |
| GPT-4oDate=2024.05, Tokenization Paradigm=Proprietary Models2026.02 | 84.6 | |
| MAmmoTH-VL-8BModel Scale=8B2024.12 | 84 | |
| Qwen2.5VLModel Scale=7B, Date=2025.02, Tokenization Paradigm=Understanding-only Models2026.02 | 83.9 | |
| Qwen2-VL-7B-ThinkingMax Output Tokens=40K2026.03 | 83.9 | |
| InternVL-2-8BModel Scale=8B2024.12 | 83.8 | |
| SAIL-VL1.5Number of Parameters=2.5B2025.12 | 83.7 | |
| Qwen2-VL-7B-ThinkingMax Output Tokens=4K2026.03 | 83.5 | |
| Qwen2.5-VL-7B + Cont. Rewardreward_type=Continuous2025.11 | 83.41 | |
| MAmmoTH-VL-8BModel Scale=8B, Single-image variant (SI)=true2024.12 | 83.4 | |
| HyperVL ViTLNumber of Parameters=2.0B2025.12 | 83.1 | |
| Qwen2-VL-7B-InsModel Scale=7B2024.12 | 83 | |
| SAIL-VL2Number of Parameters=2.7B2025.12 | 83 | |
| Ming-UnivisionModel Scale=16B, Date=2025.10, Tokenization Paradigm=Continuous Tokenization2026.02 | 82.8 | |
| Vision-Zeroexternal supervision=true2025.11 | 82.64 | |
| Qwen2.5-VL-7B (Baseline)2025.11 | 82.61 | |
| InternVL3.5Model Scale=4B, Date=2025.08, Tokenization Paradigm=Understanding-only Models2026.02 | 82.6 | |
| KelixModel Scale=8B, Tokenization Paradigm=Discrete Tokenization2026.02 | 82.4 | |
| Qwen2.5-VL-7B + Discrete Rewardreward_type=Discrete2025.11 | 82.18 | |
| MiniCPM-V-2.6-7BModel Scale=7B2024.12 | 82.1 | |
| HyperVLNumber of Parameters=1.8B2025.12 | 81.8 | |
| Circle-RoPEParameters=3B, Supervised Fine-Tuning (SFT)=true2025.05 | 81.8 | |
| LLaVA-OV-7BModel Scale=7B, Single-image variant (SI)=true2024.12 | 81.6 | |
| Qwen2.5VLModel Scale=3B, Date=2025.02, Tokenization Paradigm=Understanding-only Models2026.02 | 81.6 | |
| Qwen2.5-VL-3B-InstructParams=3.75B2025.12 | 81.51 | |
| INXComp-2.5-7BModel Scale=2.5B2024.12 | 81.5 | |
| Zamba2-VL-1.2BParameter Scale=1.2B2026.05 | 81.5 | |
| LLaVA-OV-7BModel Scale=7B2024.12 | 81.4 | |
| LLaVA-OVModel Scale=7B, Date=2024.08, Tokenization Paradigm=Understanding-only Models2026.02 | 81.4 | |
| InternVL2.5Parameters=4B2025.05 | 81.38 | |
| MiniCPM V-2.6Parameters=8B2025.05 | 81.28 | |
| Kimi-VL-A3B-Thinking2026.03 | 81.2 | |
| InternVL3.5-1BParameter Scale=1B2026.05 | 81.2 | |
| Molmo-7B-DModel Scale=7B2024.12 | 81 | |
| InternVL 1.5#param=26B, open-source=false2024.04 | 80.7 | |
| gemma-2-9b-it2026.03 | 80.4 | |
| Gemini Pro 1.5open-source=false2024.04 | 80.3 | |
| Claude3.5-Sonnet2024.10 | 80.2 | |
| AndesVLNumber of Parameters=2.4B2025.12 | 80.1 | |
| Ovis2Number of Parameters=2.5B2025.12 | 79.7 | |
| Phi-4-reasoning-vision-15BThinking Mode=force thinking2026.03 | 79.7 | |
| Gemini Ultra 1.0open-source=false2024.04 | 79.5 | |
| Qwen-VL-Maxopen-source=false2024.04 | 79.3 | |
| TUNAModel Scale=7B, Date=2025.12, Tokenization Paradigm=Continuous Tokenization2026.02 | 79.3 | |
| Qwen2.5-VL (SFT)Parameters=3B, Supervised Fine-Tuning (SFT)=true2025.05 | 79.22 | |
| Step-1V#param=100B, open-source=false2024.04 | 79.2 | |
| Gemini-1.5-Pro2024.10 | 79.1 | |
| InternVL 1.2#param=40B, open-source=false2024.04 | 79 | |
| InternVL3.5Number of Parameters=2.3B2025.12 | 78.8 | |
| InternVL3Number of Parameters=2.1B2025.12 | 78.7 | |
| MobileNet-QwenParams=1.84B2025.12 | 78.63 | |
| Show-O2Model Scale=7B, Date=2025.06, Tokenization Paradigm=Discrete Tokenization2026.02 | 78.6 | |
| Phi-3-VisionParams (B)=4.22024.10 | 78.4 | |
| GPT-4Vopen-source=false2024.04 | 78.2 | |
| Qwen2.5-VLParameters=3B2025.05 | 78.14 | |
| SAIL-VLParameters=2B2025.05 | 77.72 | |
| Phi-3.5 visionParameters=4.2B2025.05 | 77.59 | |
| Llama-3.2-11B-Vision-InsModel Scale=11B2024.12 | 77.3 | |
| InternViT-QwenParams=1.86B2025.12 | 77.17 | |
| Qwen3-VLNumber of Parameters=2.1B2025.12 | 76.9 | |
| X-OmniModel Scale=7B, Date=2025.07, Tokenization Paradigm=Discrete Tokenization2026.02 | 76.8 | |
| Qwen-VL-Plusopen-source=false2024.04 | 75.9 | |
| TokenFlow-XLModel Scale=14B, Date=2024.12, Tokenization Paradigm=Discrete Tokenization2026.02 | 75.8 | |
| LLaVA-NeXT#param=35B, open-source=false2024.04 | 74.9 | |
| Qwen2-VL-2BParams (B)=2.12024.10 | 74.7 | |
| Qwen2-VLNumber of Parameters=2.2B2025.12 | 74.7 | |
| Qwen2-VL-2BParams=2.2B2025.12 | 74.64 |