Qianfan OCR vs Florence-2
Pick Qianfan OCR for full-page document parsing: it outputs Markdown, JSON, HTML and LaTeX, covers 192 languages and handles handwriting, tables and formulas. Pick Florence-2 for small, English scene text or short labels where you need boxes and a 0.23B or 0.77B model. The main split is structure and language: Florence-2 returns only text and boxes, and its Chinese output is mostly wrong.
Qianfan OCRbaidu/qianfan-ocr | Florence-2microsoft/florence-2 | |
|---|---|---|
| Type | OCR API | Open source |
| Price | 2 credits / page | 1 credit / page |
| Per 1,000 pages | $2.00 | $1.00 |
| Typical speed | 2.0 s / page | 0.8 s / page |
| Best for | Chinese documents | Lightweight OCR with boxes |
| Inputs | PDF, images | images |
| Languages | Chinese, English and more | English |
| Max pages per request | 200 | 500 |
| Tables | ✓ Yes | No |
| Handwriting | ✓ Yes | No |
| Math | ✓ Yes | No |
| Bounding boxes | No | ✓ 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 Qianfan OCR if
- You need structured output such as Markdown, JSON, HTML, LaTeX or bounding boxes from full documents.
- You parse multilingual files, handwriting or scene text, since it covers 192 languages.
- You want a measured document score: 93.12 on OmniDocBench v1.5, plus 79.8 on olmOCR-Bench and 880 on OCRBench.
Choose Florence-2 if
- You need a very small model: 0.23B base or 0.77B large, MIT licensed.
- You read English scene text or short labels and want quadrilateral boxes per text line from OCR_WITH_REGION.
- Your pipeline also needs captioning, detection or grounding from the same model.
Key differences
| Qianfan OCR | Florence-2 | |
|---|---|---|
| Model size | 4B (Qwen3-4B language model) | 0.23B (base), 0.77B (large) |
| Licence | Apache-2.0, allows commercial use | MIT |
| How you run it | Self-host the open weights, or use the Baidu Qianfan platform | Self-host via Transformers only |
| Output formats | Markdown, JSON, HTML, LaTeX, bounding boxes | Plain text and bounding boxes. No Markdown, tables or LaTeX |
| Languages and handwriting | 192 languages, handwriting and scene text | Chinese output mostly wrong because the vocabulary lacks CJK characters. Spelling errors and missing spaces on cursive handwriting and French text in one test |
| Benchmarks | OmniDocBench v1.5 93.12, olmOCR-Bench 79.8, OCRBench 880 | TextCaps CIDEr zero-shot 72.8 (large), COCO detection mAP zero-shot 37.5 (large). No document OCR benchmark published |
| Known limits | 1.024 pages per second on an A100 even with W8A8. Thinking mode can emit up to 16,384 tokens. Not suited to CPU-only deployment | Model card notes the continued-pretrained checkpoint used only 0.1B samples and might not be well trained |
Sources
- Qianfan-OCR model card
- Qianfan-OCR paper
- Florence-2-large model card
- HF blog OCR test
- HF discussion Chinese very bad
What each one is
Qianfan OCR
Baidu's OCR-tuned vision model. Strong on Chinese documents and tables.
Qianfan 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": "baidu/qianfan-ocr","fallbacks": ["microsoft/florence-2"],"file_url": "https://example.com/document.pdf"}Questions
More comparisons
Qianfan OCR alternativesFlorence-2 alternativesBest OCR models in 2026