Optical Character Recognition on OCRBench v2
68Accuracy (en)Gemini-3-Pro
Evaluation Results
| Method | Links | |||
|---|---|---|---|---|
| Gemini-3-Pro2026.03 | 68 | 52.5 | — | |
| Qwen3-VL-235B-ThinkingNumber of Parameters=235B-A22B, Thinking Configuration=true2026.03 | 66.8 | 63.8 | — | |
| Kimi-K2.5Number of Parameters=1T-A32B2026.03 | 64.2 | 57.4 | — | |
| Qwen3-VL-4BParameters=4B2026.03 | 60.68 | 59.13 | — | |
| Intern-S1-ProNumber of Parameters=1T-A22B2026.03 | 60.1 | 60.6 | — | |
| GPT-5.22026.03 | 56.4 | 54.6 | — | |
| Qianfan-OCR2026.03 | 56 | 60.77 | — | |
| Qwen3-VL-2BParameters=2B2026.03 | 54.18 | 54.41 | — | |
| PointsReader2026.03 | 23.4 | 40.39 | — | |
| MonkeyOCR2026.03 | 21.78 | 38.91 | — | |
| Dotsocr2026.03 | 19.59 | 35.94 | — | |
| PaddleOCR-VL2026.03 | 18.15 | 40.86 | — | |
| MinerU2.52026.03 | 17.09 | 40.03 | — | |
| DeepSeek-OCR2026.03 | 15.98 | 38.31 | — | |
| EXAONE 4.5 33BMode=REASONING, Architecture=Dense, # Total Params=33B, # Activated Params=33B2026.04 | — | — | 63.2 | |
| FullTraining Strategy=Full-Attention baseline2026.06 | — | — | 55 | |
| GPT-5 miniMode=REASONING: HIGH, Architecture=-, # Total Params=-, # Activated Params=-2026.04 | — | — | 55.8 | |
| MSA-CPTTraining Strategy=sparse continued pretraining2026.06 | — | — | 54.3 | |
| MSA-PTTraining Strategy=from-scratch sparse pretraining2026.06 | — | — | 55.7 | |
| Qwen3-VL-235B-A22BMode=Thinking, Architecture=MoE, # Total Params=236B, # Activated Params=23B2026.04 | — | — | 66.8 | |
| Qwen3-VL-32BMode=Thinking, Architecture=Dense, # Total Params=33B, # Activated Params=33B2026.04 | — | — | 68.4 | |
| Qwen3.5-27BMode=REASONING, Architecture=Dense, # Total Params=27B, # Activated Params=27B2026.04 | — | — | 67.3 |