dots.ocr vs Florence-2
Pick dots.ocr when you need full documents turned into Markdown, HTML tables or LaTeX, especially multilingual PDFs. Pick Florence-2 when you need small, fast text with line boxes from scene text or short labels in English, or when your pipeline also needs detection. The main split is structure: dots.ocr returns layout, tables and formulas, while Florence-2 returns only text and boxes.
dots.ocrrednote/dots-ocr | Florence-2microsoft/florence-2 | |
|---|---|---|
| Type | Open source | Open source |
| Price | 2 credits / page | 1 credit / page |
| Per 1,000 pages | $2.00 | $1.00 |
| Typical speed | 1.5 s / page | 0.8 s / page |
| Best for | Multilingual document parsing | Lightweight OCR with boxes |
| Inputs | PDF, images | images |
| Languages | 100+ languages | English |
| Max pages per request | 500 | 500 |
| Tables | ✓ Yes | No |
| Handwriting | No | No |
| Math | ✓ Yes | No |
| 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 dots.ocr if
- It outputs layout boxes, categories and text in one pass, with tables as HTML and formulas as LaTeX.
- It is built for multilingual PDFs, academic papers and business documents with mixed tables and text.
- It scores 79.1 on olmOCR-bench, and the dots.mocr (v1.5) version reaches 83.9.
Choose Florence-2 if
- It is tiny at 0.23B (base) or 0.77B (large) parameters, against 3B for dots.ocr.
- OCR_WITH_REGION returns a quadrilateral box for each text line, which suits scene text and short labels.
- The same model also handles captioning, detection and grounding, so one model can cover several steps of a pipeline.
Key differences
| dots.ocr | Florence-2 | |
|---|---|---|
| Output formats | JSON, Markdown, HTML, LaTeX, bounding boxes, SVG | Plain text and bounding boxes only |
| Model size | 3B (1.7B LLM backbone) | 0.23B (base), 0.77B (large) |
| Licence | MIT (the repository also has a separate dots.ocr licence agreement file) | MIT |
| How you run it | Self-hosted via vLLM or Hugging Face Transformers. A single CUDA GPU with vLLM is the documented path. No VRAM figure is stated. | Self-hosted via Transformers. Pitched for lightweight edge use. |
| Tables and formulas | Returns tables as HTML and formulas as LaTeX. Authors say complex tables and formulas are still hard at this size, and a table above 1000 grid cells raised an error and dropped the whole page in a Docling issue. | No table, Markdown or LaTeX output. |
| Chinese and CJK text | Described as multilingual. CJK accuracy is Not published. | Chinese output was mostly wrong, because the vocabulary lacks CJK characters. |
| Known failure | Runs of ellipses or underscores can make the output repeat endlessly, so forms with long dot leaders are a risk. | A Hugging Face blog test found spelling errors and missing spaces on cursive handwriting and French text. |
Sources
- dots.ocr GitHub README
- dots.ocr blog / limitations
- Docling issue #5146
- Florence-2-large model card
- HF blog OCR test
- HF discussion Chinese very bad
What each one is
dots.ocr
Single vision model from rednote that does layout detection and content recognition together.
dots.ocr detailsFlorence-2
Small Microsoft vision model (0.23B and 0.77B) with OCR and OCR-with-regions tasks that return text with boxes.
Florence-2 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": "rednote/dots-ocr","fallbacks": ["microsoft/florence-2"],"file_url": "https://example.com/document.pdf"}Questions
More comparisons
dots.ocr alternativesFlorence-2 alternativesBest OCR models in 2026