Updated 2026-11-02
Best Arabic OCR models in 2026, compared
Scores from the two public Arabic OCR benchmarks, KITAB-Bench and Misraj-DocOCR, for open models, cloud APIs and classic OCR engines. Other systems' numbers are the ones published by the benchmark authors.
Summary
- Most accurate overall: Nuqta-Arabic-OCR 4B (open, Apache-2.0). First on both benchmarks.
- Best small model: Nuqta-Arabic-OCR 0.8B. Second on KITAB-Bench, ahead of Gemini-2.0-Flash and GPT-4o, and runs on a phone.
- Lowest character error on Misraj-DocOCR: Baseer v2 (0.12) and Nuqta 4B (0.121) are effectively tied; Baseer v2 has no public weights.
- Classic engines: Tesseract (0.54) and EasyOCR (0.58) miss about half of the characters on KITAB-Bench.
KITAB-Bench: 13 Arabic OCR datasets
3,760 images: book pages, scanned lines, handwriting, manuscripts, slides, scene text and fonts. Character error rate, lower is better.
| Model | Weights | CER ↓ |
|---|---|---|
| Nuqta-Arabic-OCR 4B | 4.75B, open | 0.106 |
| Nuqta-Arabic-OCR 0.8B | 0.89B, open | 0.117 |
| Gemini-2.0-Flash | closed | 0.130 |
| AIN-7B | 7B, open | 0.200 |
| GPT-4o | closed | 0.310 |
| Qwen-2.5-VL | 7B, open | 0.492 |
| Tesseract | open | 0.540 |
| EasyOCR | open | 0.580 |
Nuqta at its recommended input size (small page renders are enlarged to the training size). With every image sent unchanged: 4B 0.111, 0.8B 0.132. 234 test images also appear in public datasets used for training; without them the 4B scores 0.113. Source: KITAB-Bench and the Nuqta model cards.
Misraj-DocOCR: 400 real Arabic document pages
Pages converted to Markdown. MARS combines text and table accuracy, TEDS scores tables; higher is better. CER, lower is better.
| Model | Weights | MARS ↑ | TEDS ↑ | CER ↓ |
|---|---|---|---|---|
| Nuqta-Arabic-OCR 4B | 4.75B, open | 82.8 | 76.1 | 0.121 |
| Baseer v2 (Misraj) | no public weights | 77.8 | 67 | 0.12 |
| Nuqta-Arabic-OCR 0.8B | 0.89B, open | 77.1 | 66.8 | 0.130 |
| Baseer (paper) | 3B | 76.9 | 66 | 0.53 |
| Gemini-2.5-Pro | closed | 70.8 | 52 | 0.31 |
| Azure Document Intelligence | closed | 62.2 | 42 | 0.27 |
| dots.ocr | 3B, open | 59.2 | 40 | 0.40 |
| GPT-5 | closed | 54.8 | 48 | 0.62 |
| Mistral OCR | closed | 44.4 | 17 | 0.42 |
| QARI | 2B, open | 42.8 | 21 | 0.64 |
Other systems from the Baseer paper and Misraj's September 2026 results table. Our scorer reproduces Misraj's published Qwen3-VL-8B score within a point, and our predictions are public in the benchmarks folder for anyone to re-score. Dataset: Misraj-DocOCR.
How to choose
- Highest accuracy, own GPUNuqta-Arabic-OCR 4B on a 16–24 GB GPU.
- Small GPU, CPU or phoneNuqta-Arabic-OCR 0.8B, or its 560 MB GGUF build.
- Coming from TesseractSame one-call use from Python. Switching guide.
Questions
Which Arabic OCR is the most accurate in 2026?
Nuqta-Arabic-OCR 4B has the best published score on both public Arabic OCR benchmarks: KITAB-Bench character error rate 0.106 and Misraj-DocOCR MARS 82.8.
Is Tesseract good for Arabic?
Tesseract works for clean printed Arabic but misses many characters on real documents: 0.54 character error rate on KITAB-Bench, against 0.106 for Nuqta-Arabic-OCR 4B. It also drops or mangles tashkeel.
Is Gemini or GPT better than open models for Arabic OCR?
Not on these benchmarks. Gemini-2.0-Flash scores 0.13 and GPT-4o 0.31 on KITAB-Bench, and Gemini-2.5-Pro 70.8 and GPT-5 54.8 MARS on Misraj-DocOCR, all behind Nuqta's open models.
What is the best Arabic OCR for tables?
On Misraj-DocOCR tables (TEDS), Nuqta-Arabic-OCR 4B scores 76.1, ahead of Baseer v2 (67), Gemini-2.5-Pro (52) and Azure (42).