Surya vs Pixtral
Pick Surya when you need layout, reading order, tables and bounding boxes from a small model you run yourself, including on a laptop. Pixtral 12B is a general vision model that did well on DocVQA, so it only fits question answering over documents and charts on an existing Mistral API setup. Mistral marks it as deprecated, so Surya is the safer choice for new OCR work.
Suryadatalab/surya | Pixtralmistral/pixtral | |
|---|---|---|
| Type | Open source | Vision model |
| Price | 2 credits / page | 2 credits / page |
| Per 1,000 pages | $2.00 | $2.00 |
| Typical speed | 1.5 s / page | 3.0 s / page |
| Best for | Layout and line-level OCR | Questions about document images |
| Inputs | PDF, images | images |
| Languages | 90+ languages | Multilingual |
| Max pages per request | 500 | 500 |
| Tables | ✓ Yes | No |
| Handwriting | No | No |
| Math | No | No |
| Bounding boxes | ✓ Yes | No |
| Confidence | No | No |
| Forms | No | 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 Surya if
- It returns layout labels, reading order, OCR, HTML tables and KaTeX math with per-block bounding boxes from one 689M model.
- It runs on CPU and Apple Silicon through llama.cpp, with peak memory of about 4-6 GB on an M4 Max.
- It scores 83.7% on olmOCR-bench (per its authors) and reached 4.52 pages/s at 192 DPI on one H100 with vLLM.
Choose Pixtral if
- It scores 90.7 ANLS on DocVQA (per Mistral), which suits question answering over documents and charts.
- Its 128K token context lets you send many images in one request.
- The Mistral API charged $0.15 per million input and output tokens at launch, and the weights are Apache-2.0, so teams already on that API can keep their integration.
Key differences
| Surya | Pixtral | |
|---|---|---|
| What it is | OCR, layout, reading order and table toolkit, 689M parameters | General vision language model, 12.4B parameters (12B decoder plus 400M vision encoder) |
| Licence | Code Apache-2.0. Weights under a modified AI Pubs Open Rail-M: free for research, personal use and companies under $5M funding or revenue | Apache-2.0 |
| How you run it | Self-host only. Needs a separate vLLM or llama.cpp server; NVIDIA GPU, CPU or Apple Silicon | Mistral API (pixtral-12b-2409) or self-hosted open weights |
| Price | Free to self-host, you pay for compute. Commercial licence needed above $5M revenue or funding | $0.15 per million input and output tokens at API launch. Free to self-host, you pay for compute |
| Output formats | JSON, HTML, LaTeX, bounding boxes, plain text | Plain text, Markdown, JSON |
| Benchmarks | olmOCR-bench 83.7%; 74.9% pass rate across 91 languages | DocVQA 90.7 ANLS; ChartQA 81.8; MMMU 52.5 |
| Known failure | Surya 2.5 table output lacks row and column geometry. Multilingual tables score 58.1. Not for handwriting or photos | Reported drops on handwriting, small fonts and non-Latin scripts. Q4 quantization is reported to hurt small text. Deprecated by Mistral |
Sources
- Surya GitHub README
- Datalab pricing and license FAQ
- Mistral announcement
- Mistral Sept 2024 pricing
- HN launch thread
What each one is
Surya
Open-source OCR toolkit with line detection, layout analysis, reading order and tables.
Surya detailsPixtral
Mistral's open-weight vision models, Pixtral 12B and Pixtral Large. Mistral's API no longer serves them, so they run self-hosted.
Pixtral 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/surya","fallbacks": ["mistral/pixtral"],"file_url": "https://example.com/document.pdf"}Questions
More comparisons
Surya alternativesPixtral alternativesBest OCR models in 2026