Marker vs Florence-2
Pick Marker when you need full PDF pages converted to Markdown, HTML or JSON, such as books, papers with equations, or RAG chunking, and your company fits the weight licence. Pick Florence-2 when you need text with bounding boxes from scene text or short English labels on small hardware, or a single model that also does detection and captioning. The main split is output: Marker returns document structure, Florence-2 returns only text and boxes.
Markerdatalab/marker | Florence-2microsoft/florence-2 | |
|---|---|---|
| Type | OCR API | Open source |
| Price | 14 credits / page | 1 credit / page |
| Per 1,000 pages | $14.00 | $1.00 |
| Typical speed | 2.5 s / page | 0.8 s / page |
| Best for | Books and research papers | Lightweight OCR with boxes |
| Inputs | PDF, images, DOCX | images |
| Languages | 90+ languages | English |
| Max pages per request | 1,000 | 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 Marker if
- You want structured output from whole documents: Markdown, HTML, JSON block trees, LaTeX or RAG chunks.
- You need a CPU option: fast mode runs on CPU with about 18 MB of ONNX weights at 23.6 pages per second, and vllm mode scores 83.7 on olmOCR-bench (Datalab's own number).
- You want a hosted fallback: Datalab's API includes about 5,000 free pages per month.
Choose Florence-2 if
- You need a quadrilateral box for each text line, which OCR_WITH_REGION returns.
- You want a small MIT-licensed model: 0.23B base or 0.77B large, with no revenue limit on commercial use.
- You also need captioning, detection or grounding from the same model.
Key differences
| Marker | Florence-2 | |
|---|---|---|
| Licence | Code Apache-2.0. Weights under a modified AI Pubs Open Rail-M licence: free for research, personal use and startups under $5M. | MIT |
| Commercial use | Companies above $5M need a paid licence for the weights. | No stated restriction beyond MIT. |
| How you run it | Local on GPU, CPU or Apple Silicon (MPS), or Datalab's hosted API. vllm mode needs an NVIDIA GPU. | Self-hosted via Transformers only. |
| Model size | Fast mode uses about 18 MB of ONNX weights on CPU. | 0.23B (base) or 0.77B (large) parameters. |
| Output | Markdown, HTML, JSON, chunks, LaTeX, images. | Plain text and bounding boxes. No Markdown, tables or LaTeX. |
| Benchmark | 83.7 on olmOCR-bench and 85.9 on OmniParseBench in vllm mode. Fast mode on CPU scores 66.6 on olmOCR-bench. All are Datalab's own numbers. | No OCR benchmark published. Reported scores are TextCaps CIDEr 72.8 and COCO detection mAP 37.5, zero-shot, large model. |
| Known failure | Not published | Spelling errors and missing spaces on cursive handwriting and French text in an HF blog test. Chinese output mostly wrong because the vocabulary lacks CJK characters. |
Sources
- Marker GitHub README
- Datalab pricing
- Florence-2-large model card
- HF blog OCR test
- HF discussion Chinese very bad
What each one is
Marker
Hosted version of the open-source Marker and Surya models. Strong on books and papers.
Marker 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": "datalab/marker","fallbacks": ["microsoft/florence-2"],"file_url": "https://example.com/document.pdf"}Questions
More comparisons
Marker alternativesFlorence-2 alternativesBest OCR models in 2026