SmolDocling vs RolmOCR
Pick RolmOCR for plain text transcription of scanned pages, especially rotated or skewed ones, on a single GPU. Pick SmolDocling only for English experiments where you need tables, formulas, code and bounding boxes from a 256M model. SmolDocling is a preview release that its authors have replaced with Granite-Docling, so it is a poor fit for production.
SmolDoclinghuggingface/smoldocling | RolmOCRreducto/rolmocr | |
|---|---|---|
| Type | Open source | Open source |
| Price | 1 credit / page | 2 credits / page |
| Per 1,000 pages | $1.00 | $2.00 |
| Typical speed | 1.0 s / page | 3.0 s / page |
| Best for | Very cheap page conversion | Fast open PDF to text |
| Inputs | PDF, images | PDF, images |
| Languages | Mainly English | Not stated by the authors |
| Max pages per request | 500 | 500 |
| Tables | ✓ Yes | ✓ Yes |
| Handwriting | No | ✓ Yes |
| Math | ✓ Yes | No |
| Bounding boxes | ✓ Yes | No |
| 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 returns structure: DocTags, Markdown, HTML, JSON, LaTeX and bounding boxes, while RolmOCR returns plain text only.
- At 256M parameters it runs on modest hardware, and it reaches about 0.35 seconds per page on an A100 with vLLM.
- One model covers text, tables (OTSL), code, formulas and charts with location tags, so you can test a full document pipeline with it.
Choose RolmOCR if
- It is built for scans: about 15% of its training data was rotated to handle off-angle pages.
- It is Apache-2.0 and not a preview, so you can use it commercially. SmolDocling's own licence information is inconsistent (Apache-2.0 in the model summary, CDLA-Permissive-2.0 in the metadata tag).
- It avoids SmolDocling's looping failure. SmolDocling can repeat tokens endlessly or drop closing tags, and Docling issue #1515 reports very slow generation on some images.
Key differences
| SmolDocling | RolmOCR | |
|---|---|---|
| Output | DocTags, Markdown, HTML, JSON, LaTeX, bounding boxes | Plain text only, no layout bounding boxes |
| Model size | 256M parameters | 8B (Qwen2.5-VL-7B base) |
| Licence | Apache-2.0 per model summary; metadata tag says CDLA-Permissive-2.0 | Apache-2.0 |
| Languages | English only | Not published |
| Release status | Preview; superseded by Granite-Docling | Released by Reducto, fine-tuned from Qwen2.5-VL-7B-Instruct on olmOCR-mix-0225 |
| Reported benchmark | Table TEDS 0.82 (FinTabNet), full-page OCR F1 0.80, OCRBench 338, all reported by the authors | 0.663 overall on the independent Needl.ai PDF extraction test, versus 0.770 for olmocr-2-7b |
| Known failure | Endless token repetition, missing closing tags, slow looping output on some images | Can hallucinate or drop content, and omitted a table subtitle in a Reducto example |
Sources
- SmolDocling model card
- Docling issue #1515
- Granite-Docling comparison table
- Hugging Face model card
- Needl.ai benchmark
What each one is
SmolDocling
Tiny 256M vision model from Hugging Face and IBM that outputs DocTags. Predecessor of Granite Docling.
SmolDocling detailsRolmOCR
Reducto's open PDF-to-text model, built on Qwen2.5-VL-7B with the olmOCR recipe. Needs no PDF metadata and handles rotated pages.
RolmOCR 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": ["reducto/rolmocr"],"file_url": "https://example.com/document.pdf"}Questions
More comparisons
SmolDocling alternativesRolmOCR alternativesBest OCR models in 2026