PaddleOCR

Open-source OCR toolkit with a dedicated Arabic model and a VLM variant

Visit website →

About

PaddleOCR is Baidu's open-source OCR toolkit, Apache 2.0 licensed, covering detection, recognition, layout analysis and table extraction with models that run on CPUs, GPUs and mobile devices. Its latest generation ships a dedicated Arabic-script recognition model covering Arabic, Persian, Urdu, Pashto, Kurdish, Uyghur, Sindhi, Balochi and English, for which the project reports 81 percent accuracy on an internal test set of about 2,700 images, a gain of more than twenty points over the previous generation. A separate vision-language model, PaddleOCR-VL, claims 109 languages with Arabic named explicitly and describes itself as suited to handwritten and historical documents, without publishing Arabic figures. The one independent Arabic number we found is less flattering: KITAB-Bench measured a character error rate of 0.79 for PaddleOCR in 2025, before the current Arabic model. Handwriting support for Arabic is not documented. It is a strong choice for teams with engineering capacity who need self-hosted Arabic OCR with tables and layout.

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

Why we picked it

It is the only major open-source toolkit that trained a model specifically for the Arabic script family and published a number for it. We list it as the self-hosted step up from Tesseract for teams who can run Python and GPUs, with the caveat that its only independent Arabic score predates the current model.

Featured in our guides