PaddleOCR-VL vs Florence-2
Pick PaddleOCR-VL for full-page document parsing: scanned reports, academic papers with formulas, tables and charts, and text in 109 languages, with Markdown or JSON output. Pick Florence-2 only for small jobs such as English scene text or short labels where you need line boxes, or where the same model must also do captioning or detection. The main split is output: Florence-2 returns plain text and boxes, with no tables, Markdown or LaTeX.
PaddleOCR-VLpaddle/paddleocr-vl | Florence-2microsoft/florence-2 | |
|---|---|---|
| Type | Open source | Open source |
| Price | 6 credits / page | 1 credit / page |
| Per 1,000 pages | $6.00 | $1.00 |
| Typical speed | 1.3 s / page | 0.8 s / page |
| Best for | CJK documents | Lightweight OCR with boxes |
| Inputs | PDF, images | images |
| Languages | 109 languages | English |
| Max pages per request | 500 | 500 |
| Tables | ✓ Yes | No |
| Handwriting | No | No |
| Math | ✓ Yes | No |
| Bounding boxes | ✓ Yes | ✓ Yes |
| Confidence | ✓ Yes | 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 PaddleOCR-VL if
- You need whole pages converted to Markdown or JSON, including tables, formulas and charts.
- You process multilingual documents, since it covers 109 languages including Cyrillic, Arabic, Devanagari and Thai.
- You want a published document benchmark result: version 1.6 reports 96.33 on OmniDocBench v1.6, and olmOCR-Bench lists 80.0 overall.
Choose Florence-2 if
- You need the smallest possible model: 0.23B base or 0.77B large, under an MIT licence.
- You need a quadrilateral box for each text line, which OCR_WITH_REGION returns.
- You want one model that also handles captioning, detection and grounding next to OCR.
Key differences
| PaddleOCR-VL | Florence-2 | |
|---|---|---|
| Output formats | Markdown and JSON for full pages | Plain text and bounding boxes only; no Markdown, tables or LaTeX |
| Model size | 0.9B parameters | 0.23B (base) or 0.77B (large) |
| Licence | Apache-2.0 | MIT |
| Languages | 109 languages in one model | Chinese output mostly wrong because the vocabulary lacks CJK characters; spelling errors and missing spaces reported on French text |
| Tables, formulas, charts | Recognizes text, tables, formulas and charts | Not supported as structured output |
| Hardware and running it | Self-host via PaddleOCR, vLLM or transformers. NVIDIA GPU recommended; CPU possible but slow. Minimum VRAM not published | Self-host via Transformers. Hardware requirements not published |
| Known failure | Version 1.6: Cyrillic regression on born-digital PDFs, missing table content, hangs on blurry images, and use_cache set to false makes decoding about 11x slower | Spelling errors and missing spaces on cursive handwriting; the model card notes the continued-pretrained checkpoint may not be well trained |
Sources
- Hugging Face model card
- PaddleOCR-VL-1.6 model card
- olmOCR-Bench table
- Issue #18453 use_cache slowdown
- Florence-2-large model card
- HF discussion Chinese very bad
What each one is
PaddleOCR-VL
Open-source document model from Baidu. Excellent on Chinese, Japanese and Korean.
PaddleOCR-VL 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": "paddle/paddleocr-vl","fallbacks": ["microsoft/florence-2"],"file_url": "https://example.com/document.pdf"}Questions
More comparisons
PaddleOCR-VL alternativesFlorence-2 alternativesBest OCR models in 2026