docTR vs SmolDocling
Pick docTR when you need words and lines with coordinates from invoices or receipts and plan to feed them into your own extraction code. Pick SmolDocling only for experiments converting English pages to Markdown, HTML or JSON with tables and formulas, and move to Granite-Docling for production. The main split is that docTR returns text boxes only, while SmolDocling returns full document structure but is a preview model with looping failures.
docTRmindee/doctr | SmolDoclinghuggingface/smoldocling | |
|---|---|---|
| Type | Open source | Open source |
| Price | 1 credit / page | 1 credit / page |
| Per 1,000 pages | $1.00 | $1.00 |
| Typical speed | 0.9 s / page | 1.0 s / page |
| Best for | Word-level OCR with positions | Very cheap page conversion |
| Inputs | PDF, images | PDF, images |
| Languages | Mainly Latin scripts | Mainly English |
| Max pages per request | 500 | 500 |
| Tables | No | ✓ Yes |
| Handwriting | No | No |
| Math | No | ✓ Yes |
| Bounding boxes | ✓ Yes | ✓ Yes |
| Confidence | ✓ Yes | 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 docTR if
- You need word-level OCR with a nested page, block, line and word structure and coordinates as JSON.
- You train or tune your own models, since you can choose the detection and recognition architectures and use straight or rotated boxes.
- You want Docker images and a hosted demo on Hugging Face Spaces, plus a key information extraction predictor.
Choose SmolDocling if
- You need Markdown, HTML, JSON or LaTeX from a page, with tables (OTSL), code, formulas and charts handled by one model.
- You want a 256M parameter model that runs on modest hardware, at about 0.35 seconds per page on an A100 with vLLM.
- You are experimenting with tiny document VLMs on English pages and a low budget.
Key differences
| docTR | SmolDocling | |
|---|---|---|
| Approach | Two-stage text detection and recognition library on PyTorch | 256M parameter vision-language model that outputs DocTags for Docling |
| Licence | Apache-2.0 | Apache-2.0 per the model summary; the metadata tag says CDLA-Permissive-2.0 |
| Output formats | JSON, plain text, bounding boxes | DocTags, Markdown, HTML, JSON, LaTeX, bounding boxes |
| Tables and formulas | No table, formula or Markdown output; words and lines only | Tables, code, formulas and charts in one model; TEDS 0.82 on FinTabNet at 150dpi (authors' figure) |
| Hardware | CPU or NVIDIA GPU; GPU Docker images need CUDA 12.2 or newer | Small enough for modest hardware; about 0.35 seconds per page on an A100 with vLLM |
| Languages | Not published | English only |
| Known failure | Split-crop merging misjoins readings across repeated characters (issue #2229); cropping images gives poorer results (issue #2007); ONNX export of MASTER truncates decoding steps (issue #2224) | Endless token repetition and missing closing tags; output loops on some images and generation becomes very slow (Docling issue #1515) |
Sources
- docTR GitHub README
- docTR documentation
- Issue #2229
- SmolDocling model card
- Docling issue #1515
- Granite-Docling comparison table
What each one is
docTR
Open-source OCR library from Mindee with separate text detection and recognition models.
docTR detailsSmolDocling
Tiny 256M vision model from Hugging Face and IBM that outputs DocTags. Predecessor of Granite Docling.
SmolDocling 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": "mindee/doctr","fallbacks": ["huggingface/smoldocling"],"file_url": "https://example.com/document.pdf"}Questions
More comparisons
docTR alternativesSmolDocling alternativesBest OCR models in 2026