Qwen-VL vs Florence-2
Pick Qwen-VL when you need OCR plus reasoning or Q&A, text in 32 languages, or Markdown, HTML and JSON output. Pick Florence-2 when you need a small MIT-licensed model that returns text with boxes for scene text and short labels in English. Qwen-VL covers far more document and language work, but it can loop and repeat text on low-quality scans.
Qwen-VLqwen/qwen-vl | 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 | 2.0 s / page | 0.8 s / page |
| Best for | Flexible extraction with prompts | Lightweight OCR with boxes |
| Inputs | PDF, images | images |
| Languages | 30+ languages | English |
| Max pages per request | 200 | 500 |
| Tables | ✓ Yes | No |
| Handwriting | ✓ Yes | 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 Qwen-VL if
- It reads text in 32 languages, while Florence-2 gets Chinese output mostly wrong because its vocabulary lacks CJK characters.
- It outputs plain text, HTML, bounding boxes, JSON and Markdown, so it can handle document structure that Florence-2 cannot.
- It comes in sizes from 2B to 235B-A22B MoE under Apache-2.0, and Alibaba Cloud also offers it as a hosted API.
Choose Florence-2 if
- At 0.23B (base) and 0.77B (large) under the MIT licence, it is far smaller and suits lightweight edge OCR in English.
- OCR_WITH_REGION returns quadrilateral boxes per text line, which suits scene text and short labels.
- The same model also does captioning, detection and grounding, so one model can cover a pipeline that needs more than OCR.
Key differences
| Qwen-VL | Florence-2 | |
|---|---|---|
| Licence | Apache-2.0 (Qwen3-VL repository) | MIT |
| How you run it | Self-host the open weights, or use the hosted API on Alibaba Cloud | Self-host via Transformers only |
| Model size | 2B, 4B, 8B, 32B dense; 30B-A3B and 235B-A22B MoE | 0.23B (base), 0.77B (large) |
| Output formats | Plain text, HTML, bounding boxes, JSON, Markdown | Plain text and bounding boxes only. No Markdown, tables or LaTeX |
| Languages | OCR in 32 languages | Chinese output mostly wrong; a test found spelling errors and missing spaces on French text |
| Handwriting | Not published | A test found spelling errors and missing spaces on cursive handwriting |
| Known failure | Can loop and repeat the same text, worse with greedy decoding and on low-quality scans. The Qwen 3 VL 8B model scores 64.6 on olmOCR-bench in a third-party table | The continued-pretrained checkpoint used only 0.1B samples and the model card says it might not be well trained |
Sources
- Qwen3-VL GitHub
- Issue #1611 infinite repetition
- Chandra README benchmark table
- Florence-2-large model card
- HF blog OCR test
- HF discussion Chinese very bad
What each one is
Qwen-VL
Alibaba's open-weight vision language models (Qwen2.5-VL, Qwen3-VL). Strong general OCR and grounding.
Qwen-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": "qwen/qwen-vl","fallbacks": ["microsoft/florence-2"],"file_url": "https://example.com/document.pdf"}Questions
More comparisons
Qwen-VL alternativesFlorence-2 alternativesBest OCR models in 2026