DeepSeek-OCR vs Florence-2
Pick DeepSeek-OCR for self-hosted bulk conversion of full pages into Markdown, HTML or text with boxes, as long as you add checks for repeated tokens. Pick Florence-2 for short English text, scene text and labels where you need boxes from a 0.23B or 0.77B model, or where the same model must also do captioning or detection. The main split is output: DeepSeek-OCR returns document formats, Florence-2 returns only text and boxes.
DeepSeek-OCRdeepseek/deepseek-ocr | Florence-2microsoft/florence-2 | |
|---|---|---|
| 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 | 0.8 s / page |
| Best for | Cheap bulk conversion | Lightweight OCR with boxes |
| Inputs | PDF, images | images |
| Languages | 100+ languages | English |
| Max pages per request | 200 | 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 DeepSeek-OCR if
- You need page-level output in Markdown, plain text, HTML or bounding boxes, and Florence-2 gives no Markdown, tables or LaTeX.
- You run bulk OCR on your own GPU and want throughput: the authors quote about 2,500 tokens per second on one A100-40G.
- You want to tune cost per page, since resolution modes range from 64 to 400 vision tokens.
Choose Florence-2 if
- You need a very small model: 0.23B base or 0.77B large, MIT licensed, for lightweight English OCR.
- You want quadrilateral boxes per text line from OCR_WITH_REGION for scene text or short labels.
- You also need captioning, detection or grounding from the same model in one pipeline.
Key differences
| DeepSeek-OCR | Florence-2 | |
|---|---|---|
| Licence | MIT, commercial use allowed | MIT |
| How you run it | Self-host with Transformers or vLLM; CUDA GPU, reference setup CUDA 11.8 and PyTorch 2.6 in bfloat16 | Self-host via Transformers; 0.23B (base) or 0.77B (large) parameters |
| Price | Free, you pay for compute | Free, you pay for compute |
| Output formats | Markdown, plain text, HTML, bounding boxes | Plain text and bounding boxes only; no tables or LaTeX |
| Scope | OCR model that compresses pages into few vision tokens | General vision model: OCR plus captioning, detection and grounding |
| Languages and handwriting | Not covered in the sources | Chinese output mostly wrong because the vocabulary lacks CJK characters; spelling errors and missing spaces on cursive handwriting and French text in the HF blog test |
| Known failure | Repeated tokens and looping over table headers (GitHub issues #12 and #16); it corrects nonsense words instead of copying them, so odd codes may change | The continued-pretrained checkpoint used only 0.1B samples, and the model card says it might not be well trained |
Sources
- DeepSeek-OCR GitHub
- Visual Merit or Linguistic Crutch paper
- Florence-2-large model card
- HF blog OCR test
- HF discussion Chinese very bad
What each one is
Florence-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": "deepseek/deepseek-ocr","fallbacks": ["microsoft/florence-2"],"file_url": "https://example.com/document.pdf"}Questions
More comparisons
DeepSeek-OCR alternativesFlorence-2 alternativesBest OCR models in 2026