OmniDocBench
Shanghai AI Laboratory
981 annotated PDF pages across 9 document categories. Tests end-to-end document parsing including text, tables, and formulas.
Benchmark Stats
SOTA History
Coming SoonComposite Score
((1-TextEditDist)*100 + TableTEDS + FormulaCDM) / 3
Higher is better
| Rank | Model | Code | Score | Paper / Source |
|---|---|---|---|---|
| 1 | paddleocr-vl End-to-end document parsing. Score = ((1-TextEditDist)*100 + TableTEDS + FormulaCDM) / 3 | 92.86 | AlphaXiv | |
| 2 | paddleocr-vl-0.9b | 92.56 | AlphaXiv | |
| 3 | mineru-2.5 | 90.67 | AlphaXiv | |
| 4 | qwen3-vl-235b | HF | 89.15 | AlphaXiv |
| 5 | monkeyocr-pro-3b | - | 88.85 | AlphaXiv |
| 6 | ocrverse-4b 4B parameter model. Text Edit: 0.058, Formula CDM: 86.91, Table TEDS: 84.55 | - | 88.56 | github-leaderboard |
| 7 | dots-ocr-3b 3B parameter model. Text Edit: 0.048, Formula CDM: 83.22, Table TEDS: 86.78 | 88.41 | github-leaderboard | |
| 8 | gemini-25-pro | - | 88.03 | AlphaXiv |
| 9 | qwen25-vl | 87.02 | AlphaXiv | |
| 10 | mistral-ocr-2512 Same as mistral-ocr-3. Model alias for mistral-ocr-2512. Text: 90.1%, Tables: 70.9%, Formula: 78.2%. | - | 79.75 | codesota-verified |
| 11 | mistral-ocr-3 INDEPENDENTLY VERIFIED by CodeSOTA. Full benchmark run on 1355 images. Text Edit: 0.099 (90.1%), Formula Edit: 0.218 (78.2%), Table TEDS: 70.9%. Reading Order: 91.6%. | - | 79.75 | codesota-verified |
| 12 | clearocr-teamquest INDEPENDENTLY VERIFIED by CodeSOTA. Traditional OCR - text only, no table/formula recognition. Text Edit: 0.154 (84.6%), Table TEDS: 0.8%, Formula Edit: 0.902. | - | 31.7 | codesota-verified |
text-edit-distance
Higher is better
| Rank | Model | Code | Score | Paper / Source |
|---|---|---|---|---|
| 1 | clearocr-teamquest Text block recognition. 84.6% accuracy. Best on research reports (95.4%), academic papers (95.0%). | - | 0.15 | codesota-verified |
| 2 | mistral-ocr-3 Text block recognition. 90.1% accuracy. Best on academic papers (97.9%), exam papers (92.8%). | - | 0.10 | codesota-verified |
Table TEDS
Tree Edit Distance Score for table recognition
Higher is better
| Rank | Model | Code | Score | Paper / Source |
|---|---|---|---|---|
| 1 | paddleocr-vl Table structure recognition score (TEDS) | 93.52 | AlphaXiv | |
| 2 | mistral-ocr-3 Table structure recognition. TEDS Structure: 75.3%. Best on exam papers (88.0%). | - | 70.88 | codesota-verified |
| 3 | clearocr-teamquest No structured table recognition. Outputs tables as plain text. | - | 0.80 | codesota-verified |
formula-edit-distance
Higher is better
| Rank | Model | Code | Score | Paper / Source |
|---|---|---|---|---|
| 1 | clearocr-teamquest No LaTeX formula recognition. Outputs formulas as plain text. | - | 0.90 | codesota-verified |
| 2 | mistral-ocr-3 Display formula recognition. 78.2% accuracy. | - | 0.22 | codesota-verified |
reading-order
Higher is better
| Rank | Model | Code | Score | Paper / Source |
|---|---|---|---|---|
| 1 | mistral-ocr-3 Reading order accuracy. 8.4% edit distance error. | - | 91.63 | codesota-verified |
| 2 | clearocr-teamquest Reading order accuracy. 14.0% edit distance error. | - | 86.04 | codesota-verified |
OCR Edit Distance
Character-level edit distance for text extraction
Lower is better
| Rank | Model | Code | Score | Paper / Source |
|---|---|---|---|---|
| 1 | gpt-4o OCR Edit Distance (lower is better). Best on English text extraction. | - | 0.02 | AlphaXiv |
Layout mAP
Mean Average Precision for layout detection
Higher is better
| Rank | Model | Code | Score | Paper / Source |
|---|---|---|---|---|
| 1 | mineru-2.5 Layout detection mAP (highest) | 97.5 | AlphaXiv |