Surya OCR

Layout-aware open OCR for 90+ languages; its own table ranks Arabic lowest

Visit website →

About

Surya is Datalab's open-source OCR toolkit, offering text recognition in more than ninety languages together with line detection, layout analysis, reading order and table recognition, and it powers the popular Marker PDF-to-Markdown converter. The code is Apache 2.0, but the model weights carry a modified OpenRAIL-M licence that is free only below a revenue and funding threshold, which matters for commercial use. For Arabic the project is unusually candid. Its own 91-language benchmark table scores Arabic at about 45 to 46 percent, among the lowest entries, and KITAB-Bench measured a character error rate of 4.95, the worst of thirteen systems tested, with outputs often longer than the ground truth. Datalab's newer vision-language model, Chandra, reports 68 percent on Arabic against 84 percent for a Gemini model in the same table and claims handwriting support without an Arabic figure. Surya is worth knowing as the engine inside tools people already use, not as an Arabic OCR choice today.

AI-assisted draft, human-reviewed before publishing — see how we choose & review tools.

Why we picked it

Surya sits inside Marker and other PDF converters that Saudi researchers already run, so its Arabic weakness travels silently into their outputs. We list it so readers recognise the engine and check the numbers before trusting a converted Arabic PDF, and because its honesty about Arabic is a model for other vendors.

Featured in our guides