Chandra vs Surya
Pick Chandra when accuracy on scanned, multilingual or table-heavy documents matters and you want Markdown for RAG. Pick Surya when you need a small model that runs on a laptop or CPU, fast layout and reading order, or bounding boxes. Chandra scores higher on olmOCR-bench (85.8 vs 83.7%), but Surya is 689M parameters and runs on Apple Silicon.
Chandradatalab/chandra | Suryadatalab/surya | |
|---|---|---|
| Type | Open source | Open source |
| Price | 2 credits / page | 2 credits / page |
| Per 1,000 pages | $2.00 | $2.00 |
| Typical speed | 2.0 s / page | 1.5 s / page |
| Best for | Complex layouts and handwriting | Layout and line-level OCR |
| Inputs | PDF, images | PDF, images |
| Languages | 40+ languages | 90+ languages |
| Max pages per request | 500 | 500 |
| Tables | ✓ Yes | ✓ Yes |
| Handwriting | ✓ Yes | No |
| Math | ✓ Yes | No |
| Bounding boxes | ✓ Yes | ✓ Yes |
| Confidence | No | No |
| Forms | ✓ Yes | No |
| Multilingual | ✓ Yes | ✓ Yes |
| 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
- Chandra 2 scores 85.8 on olmOCR-bench, ahead of Surya 2.5 at 83.7%.
- It outputs Markdown, HTML and JSON with layout chunks, which suits PDF to Markdown for RAG and documents with tables, forms and math.
- It averages 77.8% across 43 languages, and the hosted Datalab API is available from $1 per 1,000 pages if you do not want to run a GPU.
Choose Surya if
- It runs on CPU and Apple Silicon through llama.cpp, with peak memory of about 4-6 GB on an M4 Max.
- It is faster on a GPU: 4.52 pages/s on one H100 with vLLM, against about 1.44 pages/s for Chandra.
- One model returns layout labels, reading order, OCR, HTML tables and KaTeX math with per-block bounding boxes, free below $5M revenue or funding.
Key differences
| Chandra | Surya | |
|---|---|---|
| Licence | Code Apache-2.0; weights modified OpenRAIL-M, free under $2M funding/revenue | Code Apache-2.0; weights modified AI Pubs Open Rail-M, free under $5M funding/revenue |
| How you run it | Self-host on one NVIDIA GPU with vLLM, or use the hosted Datalab API | Self-host only: NVIDIA GPU with vLLM, or CPU/Apple Silicon with llama.cpp, plus a separate inference server |
| Price | Self-hosted is free within licence terms. Hosted API: $1, $3 or $7.50 per 1,000 pages, with a $20/month free tier | Free self-hosted; you pay for compute. Commercial licence needed above $5M |
| Output formats | Markdown, HTML, JSON, bounding boxes | JSON, HTML, LaTeX, bounding boxes, plain text |
| olmOCR-bench (authors' numbers) | 85.8 (Chandra 2) | 83.7% (Surya 2.5) |
| Throughput on one H100 | About 1.44 pages/s with vLLM | 4.52 pages/s at 192 DPI with vLLM |
| Known failure | Token-loop hallucinations reported on some Russian/Kazakh scans; weak on Urdu (43.2%), Sanskrit (51.1%) and old scans | Surya 2.5 tables lack row/column geometry; multilingual tables score 58.1; handwriting and complex tables are better sent to the hosted Datalab API |
Sources
What each one is
Chandra
Datalab's OCR model that keeps layout and handles forms, tables and handwriting.
Chandra detailsSurya
Open-source OCR toolkit with line detection, layout analysis, reading order and tables.
Surya detailsTry both on your own file
Run the file once with each model and compare the output.
PDF, PNG, JPG, WebP, HEIC. Up to 20 MB each, 10 files.No file?
Use one, fall back to the other
{"model": "datalab/chandra","fallbacks": ["datalab/surya"],"file_url": "https://example.com/document.pdf"}Questions
More comparisons
Chandra vs dots.ocrolmOCR vs SuryaPaddleOCR-VL vs SuryaSurya vs Adobe Acrobat OCRDocling vs ChandraSurya vs InternVL3.5LlamaParse vs ChandraDeepSeek-OCR vs Chandra
Chandra alternativesSurya alternativesBest OCR models in 2026