SmolDocling vs OpenOCR
Pick OpenOCR if you parse Chinese or English documents, especially academic PDFs with formulas, and want a small self-hosted model with a reported 90.57 on OmniDocBench v1.5. Pick SmolDocling only for English experiments with Docling's DocTags output, since it is a preview release and its authors point users to Granite-Docling. Both are free and self-hosted, so the split is maturity and scope, not price.
SmolDoclinghuggingface/smoldocling | OpenOCRfudan/openocr | |
|---|---|---|
| Type | Open source | Open source |
| Price | 1 credit / page | 1 credit / page |
| Per 1,000 pages | $1.00 | $1.00 |
| Typical speed | 1.0 s / page | 0.5 s / page |
| Best for | Very cheap page conversion | Lightweight OCR pipeline on CPU |
| Inputs | PDF, images | PDF, images |
| Languages | Mainly English | Chinese and English |
| Max pages per request | 500 | 500 |
| Tables | ✓ Yes | ✓ Yes |
| Handwriting | No | No |
| Math | ✓ Yes | ✓ Yes |
| Bounding boxes | ✓ Yes | ✓ Yes |
| Confidence | No | No |
| Forms | No | No |
| Multilingual | No | No |
| On OCRRouter | Available | Available |
Prices are what you pay on OCRRouter. Speed is the provider's typical figure, not a benchmark we ran.
Which to choose
Choose SmolDocling if
- It outputs DocTags, Markdown, HTML and JSON, so it fits a Docling pipeline directly.
- One 256M model handles text, tables, code, formulas and charts with location tags, with no separate layout model.
- It runs at about 0.35 seconds per page on an A100 with vLLM, a published throughput figure.
Choose OpenOCR if
- OpenDoc-0.1B reports 90.57 on OmniDocBench v1.5 with 0.1B parameters.
- It supports Chinese and English, where SmolDocling is English only.
- UniRec-0.1B reads plain text and single or multi-line formulas in one model, and ONNX export is supported for deployment.
Key differences
| SmolDocling | OpenOCR | |
|---|---|---|
| Licence | Apache-2.0 per the model summary (metadata tag says CDLA-Permissive-2.0) | Apache-2.0 |
| How you run it | Self-hosted, you pay for compute | Self-hosted as a Python library, you pay for compute |
| Model size | 256M parameters | 0.1B (OpenDoc-0.1B and UniRec-0.1B) |
| Output formats | DocTags, Markdown, HTML, JSON, LaTeX, bounding boxes | Plain text, LaTeX, bounding boxes |
| Languages | English only | Chinese and English only |
| Reported benchmarks | Table TEDS 0.82 (FinTabNet 150dpi), full-page OCR F1 0.80, OCRBench 338 | OmniDocBench v1.5 overall 90.57, self-reported by the authors with no independent replication found |
| Known failure | Output can loop on some images and become very slow (Docling issue #1515); paper lists endless token repetition and missing closing tags | Document parsing depends on a separate layout model, PP-DocLayoutV2 from PaddlePaddle; no hardware or throughput figures published |
Sources
- SmolDocling model card
- SmolDocling paper
- Docling issue #1515
- Granite-Docling comparison table
- OpenOCR GitHub README
What each one is
SmolDocling
Tiny 256M vision model from Hugging Face and IBM that outputs DocTags. Predecessor of Granite Docling.
SmolDocling detailsOpenOCR
Fudan's open toolkit for text detection and recognition, formulas, tables and document parsing.
OpenOCR detailsTry both on your own file
Run the file once with each model and compare the output.
Use one, fall back to the other
{"model": "huggingface/smoldocling","fallbacks": ["fudan/openocr"],"file_url": "https://example.com/document.pdf"}Questions
More comparisons
SmolDocling alternativesOpenOCR alternativesBest OCR models in 2026