PaddleOCR-VL vs Chandra
Pick PaddleOCR-VL if you want a small Apache-2.0 model that you self-host on a modest GPU and that covers 109 languages. Pick Chandra if you need the best published olmOCR-bench score, bounding boxes in the output, or a hosted API at $1 to $7.50 per 1,000 pages. The main split is licence and deployment: PaddleOCR-VL is self-host only and free for any company, while Chandra's weights need a paid licence above $2M funding or revenue.
PaddleOCR-VLpaddle/paddleocr-vl | Chandradatalab/chandra | |
|---|---|---|
| Type | Open source | Open source |
| Price | 6 credits / page | 2 credits / page |
| Per 1,000 pages | $6.00 | $2.00 |
| Typical speed | 1.3 s / page | 2.0 s / page |
| Best for | CJK documents | Complex layouts and handwriting |
| Inputs | PDF, images | PDF, images |
| Languages | 109 languages | 40+ languages |
| Max pages per request | 500 | 500 |
| Tables | ✓ Yes | ✓ Yes |
| Handwriting | No | ✓ Yes |
| Math | ✓ Yes | ✓ Yes |
| Bounding boxes | ✓ Yes | ✓ Yes |
| Confidence | ✓ Yes | 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 PaddleOCR-VL if
- It is Apache-2.0 with no revenue cap, so large companies can self-host it commercially for free.
- At 0.9B parameters it runs on much smaller GPUs than 7B document models.
- It supports 109 languages, including Cyrillic, Arabic, Devanagari and Thai scripts, and parses text, tables, formulas and charts.
Choose Chandra if
- It scores 85.8 on olmOCR-bench, against 80.0 for PaddleOCR-VL on the same benchmark.
- It returns Markdown, HTML and JSON with bounding boxes and labels per block, and you can also call it through the hosted Datalab API.
- It reaches about 1.44 pages/s on one H100 with vLLM at 96 concurrent sequences, with a 0% failure rate.
Key differences
| PaddleOCR-VL | Chandra | |
|---|---|---|
| Licence | Apache-2.0 | Code Apache-2.0; weights modified OpenRAIL-M, free for research, personal use and companies under $2M funding/revenue |
| How you run it | Self-host only, via PaddleOCR, vLLM or transformers | Self-host with vLLM or Hugging Face, or the hosted Datalab API |
| Price | Free, you pay for compute | Self-hosted free within licence terms; API $1 (fast), $3 (balanced) or $7.50 (accurate) per 1,000 pages; free tier of $20/month |
| Size and hardware | 0.9B parameters; NVIDIA GPU recommended, CPU possible but slow; no minimum VRAM stated | Single NVIDIA GPU via vLLM; benchmarked on one H100 80GB, with presets for A100, L40S and RTX 4090/5090 |
| Output formats | Markdown, JSON | Markdown, HTML, JSON, bounding boxes |
| olmOCR-bench overall | 80.0 (independent table) | 85.8 for Chandra 2 (authors' table) |
| Known failure | Version 1.6 regresses Russian and Kazakh OCR on born-digital PDFs, hangs on blurry images, and ships with use_cache false, which makes decoding about 11x slower | Token-loop hallucinations on Russian/Kazakh scans, weak on Urdu (43.2%) and Sanskrit (51.1%), and only 51.1 on old scans in olmOCR-bench |
Sources
- Hugging Face model card
- olmOCR-Bench table
- Issue #18453 use_cache slowdown
- Chandra GitHub README
- Datalab pricing
- GitHub issue #51 token loops
What each one is
PaddleOCR-VL
Open-source document model from Baidu. Excellent on Chinese, Japanese and Korean.
PaddleOCR-VL detailsChandra
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": "paddle/paddleocr-vl","fallbacks": ["datalab/chandra"],"file_url": "https://example.com/document.pdf"}Questions
More comparisons
PaddleOCR-VL alternativesChandra alternativesBest OCR models in 2026