Category ranking · 7 models · checked 2026-10-13
OCR and documents
Ranked by olmOCR-Bench, overall. Source: allenai/olmocr README benchmark table.
olmOCR-Bench, overall
4 with a published score, 3 not yet measured.
# | Model | Parameters | Score | Source | Licence |
|---|---|---|---|---|---|
| 1 | olmOCR 2 7B Ai2 | 8.3 B | 82.4%Sourced | Source olmOCR v0.4.0 row (±1.1). | Apache-2.0 |
| 2 | PaddleOCR-VL Baidu PaddlePaddle | 959 M | 80%Sourced | Source PaddleOCR-VL* row (±1.0); asterisk as marked in the source table. | Apache-2.0 |
| 3 | DeepSeek-OCR DeepSeek | 3.3 B | 75.7%Sourced | Source DeepSeek-OCR row (±1.0). | MIT |
| 4 | Nanonets OCR2 3B Nanonets | 3.8 B | 69.5%Sourced | Source Nanonets-OCR2-3B row (±1.1). | Not declared on Hugging Face Check terms before commercial use |
| — | dots.ocr rednote hilab | 3.0 B | Not yet measured | Reports OmniDocBench, not olmOCR-Bench. | MIT |
| — | Granite Docling 258M IBM | 258 M | Not yet measured | — | Apache-2.0 |
| — | SmolDocling 256M IBM Research / Hugging Face | 256 M | Not yet measured | — | CDLA-Permissive-2.0 |
How to read this: one public benchmark per category, so you can compare like with like. We copy the published number and link it; we do not re-run it, average different benchmarks or fill gaps. Vendor-reported numbers use each vendor's own settings, so treat small gaps as ties. “Not yet measured” means we found no comparable public score, not that the model is weak.
Every model was checked against the Hugging Face API on 2026-10-13: it exists, is public, and has the parameter count and licence shown. Other categories: Math and reasoning · Coding · Vision-language · Speech to text · Text to speech · Image generation · Embeddings and rerankers · Small fast 'System 1' · Biomedical · Tool use and agents · Multilingual and translation · Time series · Robotics (VLA).
Machine-readable: /api/v1/categories/ocr-documents.json