Chandra vs Florence-2
Pick Chandra for full-page document conversion: it outputs Markdown, HTML and JSON with layout, and it covers many languages. Pick Florence-2 for short English text with boxes, or when you want a 0.23B to 0.77B MIT-licensed model that also does captioning and detection. Florence-2 returns only plain text and boxes, so it cannot give you tables or document structure.
Chandradatalab/chandra | 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 | Complex layouts and handwriting | Lightweight OCR with boxes |
| Inputs | PDF, images | images |
| Languages | 40+ languages | English |
| Max pages per request | 500 | 500 |
| Tables | ✓ Yes | No |
| Handwriting | ✓ Yes | No |
| Math | ✓ Yes | No |
| Bounding boxes | ✓ Yes | ✓ Yes |
| Confidence | No | No |
| Forms | ✓ Yes | 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 Chandra if
- You need Markdown, HTML or JSON with per-block bounding boxes and labels for scanned PDFs, tables, forms or math.
- You work with many languages: Chandra 2 averages 77.8% across 43 languages on the authors' multilingual benchmark.
- You want a hosted option: the Datalab API charges $1, $3 or $7.50 per 1,000 pages, with a $20/month free tier.
Choose Florence-2 if
- You need a small model: 0.23B (base) or 0.77B (large) parameters, against a Chandra setup that expects a single GPU such as an H100, A100, L40S or RTX 4090.
- You want an MIT licence with no revenue or funding cap and no restriction on competing with a vendor API.
- You need OCR with quadrilateral boxes per text line in the same model that also does captioning, detection and grounding.
Key differences
| Chandra | Florence-2 | |
|---|---|---|
| Licence | Code Apache-2.0; weights modified OpenRAIL-M, free under $2M funding/revenue, commercial licence needed above that | MIT |
| How you run it | Self-host via vLLM or Hugging Face, or use the hosted Datalab API | Self-host via Transformers only |
| Price | Self-hosted: free within licence terms. Datalab API: $1, $3 or $7.50 per 1,000 pages | Free, self-hosted (you pay for compute) |
| Output formats | Markdown, HTML, JSON, bounding boxes | Plain text and bounding boxes. No Markdown, tables or LaTeX |
| Hardware and size | Single NVIDIA GPU via vLLM; benchmarked on one H100 80GB | 0.23B (base) and 0.77B (large) parameters |
| Languages | 43-language average of 77.8%; weak on Urdu (43.2%) and Sanskrit (51.1%) | Chinese output mostly wrong because the vocabulary lacks CJK characters |
| Known failure | Token-loop hallucinations reported on some Russian/Kazakh scans; Datalab says raw output hallucinates on about 6 in 10 of its hardest pages before the hosted pipeline's checks | Spelling errors and missing spaces on cursive handwriting and French text |
Sources
- Chandra GitHub README
- Datalab pricing
- GitHub issue #51 token loops
- Florence-2-large model card
- HF discussion Chinese very bad
- HF blog OCR test
What each one is
Chandra
Datalab's OCR model that keeps layout and handles forms, tables and handwriting.
Chandra 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/chandra","fallbacks": ["microsoft/florence-2"],"file_url": "https://example.com/document.pdf"}Questions
More comparisons
Chandra alternativesFlorence-2 alternativesBest OCR models in 2026