DeepSeek-OCR vs SmolDocling
Pick DeepSeek-OCR for cheap self-hosted bulk OCR where an MIT licence and GPU throughput matter. Pick SmolDocling only for English page-conversion experiments on small hardware, since it is a preview release and its authors point users to Granite-Docling. Both loop on some inputs, so check output on your own tables.
DeepSeek-OCRdeepseek/deepseek-ocr | SmolDoclinghuggingface/smoldocling | |
|---|---|---|
| Type | Open source | Open source |
| Price | 1 credit / page | 1 credit / page |
| Per 1,000 pages | $1.00 | $1.00 |
| Typical speed | 1.1 s / page | 1.0 s / page |
| Best for | Cheap bulk conversion | Very cheap page conversion |
| Inputs | PDF, images | PDF, images |
| Languages | 100+ languages | Mainly English |
| Max pages per request | 200 | 500 |
| Tables | ✓ Yes | ✓ Yes |
| Handwriting | No | No |
| Math | ✓ Yes | ✓ Yes |
| Bounding boxes | ✓ Yes | ✓ Yes |
| Confidence | No | No |
| Forms | No | No |
| Multilingual | ✓ Yes | 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 DeepSeek-OCR if
- The MIT licence allows commercial use without a licence conflict.
- It offers resolution modes from 64 to 400 vision tokens per page, so you can trade detail for cost.
- The authors quote about 2,500 tokens per second of PDF throughput on one A100-40G.
Choose SmolDocling if
- At 256M parameters it is small enough to run on modest hardware.
- It outputs DocTags, Markdown, HTML, JSON and LaTeX, and one model handles text, tables, code, formulas and charts with location tags.
- The authors report about 0.35 seconds per page on an A100 with vLLM.
Key differences
| DeepSeek-OCR | SmolDocling | |
|---|---|---|
| Licence | MIT | Apache-2.0 per the model summary; the metadata tag says CDLA-Permissive-2.0 |
| How you run it | Self-host with Transformers or vLLM on a CUDA GPU (reference setup CUDA 11.8, PyTorch 2.6, bfloat16) | Self-host, 256M parameters, outputs DocTags for Docling |
| Price | Free, you pay for compute | Free, you pay for compute |
| Output formats | Markdown, plain text, HTML, bounding boxes | DocTags, Markdown, HTML, JSON, LaTeX, bounding boxes |
| Reported benchmarks | None | Table TEDS 0.82 on FinTabNet 150dpi, full-page OCR F1 0.80, OCRBench 338 (authors) |
| Languages | Not stated | English only |
| Known failures | Repeated tokens and looping over table headers (GitHub issues #12 and #16); a blank image took 30 seconds and gave repeated garbage; it corrects nonsense words instead of copying them | Endless token repetition and missing closing tags; output loops on some images and generation gets very slow (Docling issue #1515) |
| Project status | Current release | Preview release, superseded by Granite-Docling, which scores higher on every reported task (TEDS 0.97 vs 0.82) |
Sources
- DeepSeek-OCR GitHub
- Visual Merit or Linguistic Crutch paper
- LearnOpenCV test
- SmolDocling model card
- Docling issue #1515
- Granite-Docling comparison table
What each one is
SmolDocling
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.
PDF, PNG, JPG, WebP, HEIC. Up to 20 MB each, 10 files.No file?
Use one, fall back to the other
{"model": "deepseek/deepseek-ocr","fallbacks": ["huggingface/smoldocling"],"file_url": "https://example.com/document.pdf"}Questions
More comparisons
DeepSeek-OCR vs MonkeyOCRDeepSeek-OCR vs Granite DoclingJina OCR vs SmolDoclingGoogle Document AI vs SmolDoclingDeepSeek-OCR vs Tesseract.jsDeepSeek-OCR vs EasyOCRDeepSeek-OCR vs NougatDeepSeek-OCR vs Donut
DeepSeek-OCR alternativesSmolDocling alternativesBest OCR models in 2026