Nuqta OCR: open-source Arabic OCR

Accurate Arabic document OCR with faithful tashkeel, in two open models: Nuqta-Arabic-OCR 4B (#1 on KITAB-Bench and Misraj-DocOCR among published results) and Nuqta-Arabic-OCR 0.8B (best for its size: ahead of Gemini-2.0-Flash on KITAB-Bench, just behind Misraj's Baseer v2 on Misraj-DocOCR, runs on phones).

• Reads books, scans, official documents, tables, IDs and passports
• Copies the tashkeel printed on the page and almost never adds any (0.08–0.16 invented marks per 100 letters)
• Markdown output with HTML tables, in reading order
• Runs on-premises on a single GPU, or fully offline on Android and iPhone: your documents never leave your network

Install

pip install "nuqta-ocr[vllm]"   # recommended: fast, batched (Linux + NVIDIA GPU)
pip install "nuqta-ocr[fast]"   # transformers + fast kernels (no vLLM needed)
pip install nuqta-ocr           # transformers only (slow on long pages)

Quick start

from nuqta import OCR

if __name__ == "__main__":          # required with the vLLM backend (it starts worker processes)
    ocr = OCR("nuqta-0.8b")              # or "nuqta-4b" (most accurate)

    print(ocr.read("page.jpg").text)                        # as written: tashkeel only where printed
    print(ocr.read("scan.pdf", mode="markdown").text)       # headings, lists, HTML tables; every page
    print(ocr.read("passport.jpg", mode="id").fields)       # MRZ fields + check digits

    results = ocr.read_many(["a.jpg", "b.png", "c.pdf"])    # batched with vLLM

Command line:

nuqta read page.jpg
nuqta read scan.pdf --mode markdown
nuqta read passport.jpg --mode id --json

Modes

What the library does for you

• Page rule: full pages are shown to the model at its training resolution (2.4 MP); word, line and card crops keep their size. Small page renders (e.g. 72-dpi PDF pages) otherwise lose letters.
• Loop guard: if a page reading repeats itself, the page is re-read at its native size; any repetition left is trimmed (Page.looped tells you).
• PDFs: every page rendered at 200 dpi.
• ID checks: a failed check digit or an inconsistent national number is reported, never silently "fixed". Send those fields to a human.

On phones and laptops, offline

• Android and iPhone apps: apps/ has two small apps that run the 0.8B entirely on the phone (CPU on Android, GPU on iPhone), with the first words in about 2 seconds. Build steps are in apps/README.md.
• GGUF for llama.cpp, LM Studio, Ollama and other local apps: Nuqta-Arabic-OCR-0.8B-GGUF (560 MB at 4-bit) and Nuqta-Arabic-OCR-4B-GGUF (2.9 GB). The files carry the reading settings (greedy decoding, thinking off, a minimum image size), so most apps read well without setup; the cards list the settings to check in apps that override them.

Speed (0.8B, one 24 GB GPU)

Results

Recommended input size (what the library does). The 4B is first on both; the 0.8B is ahead of every other published KITAB-Bench result and third on Misraj-DocOCR, just behind Baseer v2. Our Misraj-DocOCR scorer reproduces Misraj's own published score for Qwen3-VL-8B to within a point (MARS 52.5 vs 51.8), and our exact predictions are in benchmarks/predictions for anyone to re-score. Other models' figures are the ones published by the KITAB-Bench authors and by Misraj (Baseer paper, September 2026 results table).

Full tables, confidence intervals and disclosures are in the model cards; the full benchmarks are KITAB-Bench and Misraj-DocOCR.

License

Code: Apache-2.0. Models: see their model cards.

Contributors

Salah AlHaismawi · Mohamed Saleh Tulib · Ahmed ElSayed Mohamed Ibrahim