Hey PH Arabic OCR has always been a hard problem for AI: connected script, complex ligatures, right-to-left layout, diacritics most vision-language models (including frontier ones) still struggle with it. We built Baseer to fix that: a specialized vision-language model trained on ~750K pages specifically for Arabic documents. We benchmarked it against GPT-5, Gemini 2.5 Pro, Azure Document Intelligence, Qwen2.5-VL, and DeepSeek-OCR Baseer came out ahead on every metric (WER, CER, BLEU, CHRF, TEDS). [Link to research paper] What it does beyond plain text extraction: Converts tables into clean HTML Extracts math/scientific equations as LaTeX Correctly formats classical Arabic poetry (preserving meter/rhyme structure) Detects watermarks and image placement Outputs directly to Markdown, Word, or PDF Available via API for automation, as a web app, or fully on-prem for teams that can't send documents to the cloud. Your files are never stored or used for training. We benchmarked it against GPT-5, Gemini 2.5 Pro, Azure Document Intelligence, Qwen2.5-VL, and DeepSeek-OCR — Baseer came out ahead on every metric (WER, CER, BLEU, CHRF, TEDS). Full research paper: https://arxiv.org/abs/2509.18174 Would love feedback, especially from anyone who's tried building OCR/document-AI for low-resource or complex-script languages. AMA!

Baseer

A vision-language model that outperforms GPT-5 on Arabic OCR

A vision-language model that outperforms GPT-5 on Arabic OCR