DeepSeek-OCR vs Chandra
Pick DeepSeek-OCR if you want a free, MIT-licensed model you host yourself for cheap bulk OCR, and you can add safeguards against repeated tokens. Pick Chandra for multilingual scans, tables, forms and math, or if you want a hosted API, as long as your company is under $2M in funding or revenue or you will buy a licence. The main split is licence and reach: MIT and self-host only versus modified OpenRAIL-M weights with an optional paid API.
DeepSeek-OCRdeepseek/deepseek-ocr | Chandradatalab/chandra | |
|---|---|---|
| Type | Open source | Open source |
| Price | 1 credit / page | 2 credits / page |
| Per 1,000 pages | $1.00 | $2.00 |
| Typical speed | 1.1 s / page | 2.0 s / page |
| Best for | Cheap bulk conversion | Complex layouts and handwriting |
| Inputs | PDF, images | PDF, images |
| Languages | 100+ languages | 40+ languages |
| Max pages per request | 200 | 500 |
| Tables | ✓ Yes | ✓ Yes |
| Handwriting | No | ✓ Yes |
| Math | ✓ Yes | ✓ Yes |
| Bounding boxes | ✓ Yes | ✓ Yes |
| Confidence | No | No |
| Forms | No | ✓ Yes |
| 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 DeepSeek-OCR if
- The MIT licence allows commercial use with no revenue cap, so a large company can self-host it for free.
- It offers resolution modes from 64 to 400 vision tokens per page, and the authors quote about 2,500 tokens per second PDF throughput on one A100-40G.
- It suits cheap self-hosted bulk OCR and research on token-efficient OCR, and you only pay for compute.
Choose Chandra if
- It has a reported 85.8 on olmOCR-bench (Chandra 2, authors' table), above dots.ocr 1.5 at 83.9 and olmOCR 2 at 82.4.
- It averages 77.8% across 43 languages versus 67.6% for Gemini 2.5 Flash, and suits multilingual scanned documents.
- You can skip running GPUs and use the Datalab API at $1, $3 or $7.50 per 1,000 pages, or output Markdown, HTML or JSON with layout chunks and bounding boxes.
Key differences
| DeepSeek-OCR | Chandra | |
|---|---|---|
| Licence | MIT, commercial use allowed. | Code Apache-2.0; weights modified OpenRAIL-M. Free for research, personal use and companies under $2M funding/revenue; above that you need a commercial licence. |
| How you run it | Self-host only, with Transformers or vLLM. | Self-host with vLLM or Hugging Face, or use the hosted Datalab API. |
| Price | Free, you pay for compute. | Self-hosted is free within licence terms. Datalab API is $1 (fast), $3 (balanced) or $7.50 (accurate) per 1,000 pages; free tier gives $20/month of usage with a work email. |
| Output formats | Markdown, plain text, HTML, bounding boxes. | Markdown, HTML, JSON, bounding boxes. |
| Hardware | CUDA GPU; reference setup is CUDA 11.8 and PyTorch 2.6 in bfloat16, with throughput quoted on an A100-40G. | Single NVIDIA GPU via vLLM; benchmarks on one H100 80GB. Launcher presets for A100, L40S and RTX 4090/5090. |
| Benchmarks | Not published. Throughput of about 2,500 tokens per second on an A100-40G is from the authors. | olmOCR-bench 85.8 overall, 77.8% on 43 languages and 72.7% on 90 languages (Chandra 2, authors). About 1.44 pages/s on one H100 with vLLM, 0% failure rate. |
| Known failures | Repeated tokens and looping over table headers (GitHub issues #12 and #16). A blank image took 30 seconds and gave repeated garbage. It corrects nonsense words instead of copying them. | Token-loop hallucinations on 5-6 of about 60 Russian/Kazakh scans (issue #51). Weak on Urdu (43.2%), Sanskrit (51.1%) and old scans (51.1 on olmOCR-bench). One reviewer measured 66 seconds per page on an RTX 4090 with the Hugging Face path. |
Sources
- DeepSeek-OCR GitHub
- Visual Merit or Linguistic Crutch paper
- LearnOpenCV test
- Chandra GitHub README
- Datalab pricing
- GitHub issue #51 token loops
What each one is
Chandra
Datalab's OCR model that keeps layout and handles forms, tables and handwriting.
Chandra 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": "deepseek/deepseek-ocr","fallbacks": ["datalab/chandra"],"file_url": "https://example.com/document.pdf"}Questions
More comparisons
DeepSeek-OCR alternativesChandra alternativesBest OCR models in 2026