amad-vlm7 — Reinforcement Learning with a Verifiable Reward for Thinking Arabic OCR
The thinking Arabic OCR champion: a model that reasons before it transcribes, trained with reinforcement learning against ground-truth character accuracy.
#1 on KITAB-Bench ocr-eval · 13 datasets · 3,760 images · closed weights · September 2026
amad-vlm7 is a 7B-class vision-language model that reads before it writes. Given an image of Arabic text — printed, handwritten, historical manuscript, scene text, slides, tables — it first reasons in an explicit thinking block (line by line, re-checking the words it is unsure of), then outputs the final transcription. It is the best-scoring system we know of on KITAB-Bench, the largest public Arabic OCR benchmark, ahead of Gemini-2.0-Flash, GPT-4o and the specialist Arabic model AIN-7B.
The leaderboard
KITAB-Bench ocr-eval, all 13 datasets, unmodified benchmark metrics.
| Model | CHrF ↑ | CER ↓ | WER ↓ |
|---|---|---|---|
| amad-vlm7 (ours) | 82.69 | 0.125 | 0.262 |
| amad-vlm5 (ours, previous) | 81.05 | 0.254 | 0.362 |
| AIN-7B | 78.33 | 0.20 | 0.28 |
| Gemini-2.0-Flash | 77.95 | 0.13 | 0.32 |
| GPT-4o | 61.01 | 0.31 | 0.55 |
| Qwen2.5-VL-7B | 49.23 | 1.20 | 1.41 |
| GPT-4o-mini | 47.21 | 0.43 | 0.71 |
| EasyOCR | 45.47 | 0.58 | 0.89 |
| Tesseract | 39.62 | 0.54 | 0.84 |
| Qwen2-VL-7B | 33.94 | 1.48 | 1.55 |
| Surya | 20.61 | 4.95 | 5.61 |
| Paddle | 16.73 | 0.79 | 1.02 |
Rank: #1 CHrF, #1 CER, #1 WER. Published baselines are the benchmark's own numbers (500-token budget); amad models were scored under the same metrics with a 4,096-token budget so that the thinking block fits, and with the thinking stripped before scoring — only the final transcription counts.
What thinking looks like
A real example: a six-line calligraphy panel of Diʿbil al-Khuzāʿī's elegy for Hussein.
What the panel says:
افاطم لو خلت الحسين مجدلا وقد مات عطشانا بشط فرات اذا للطمت الخد فاطم عنده واجريت دمع العين بالوجنات افاطم قومي يا ابنة الخير واندبي نجوم سموات بارض فلاة
amad-vlm7's thinking, verbatim excerpts — it does not just see letters; it recognizes what it is reading and uses that to check the words:
Line 1: "افاطم لو خلت الحسين مجدلا" — a poetic statement referencing Fatima, daughter of the Prophet Muhammad, and Hussein, her son … Line 2: "وقد مات عطشانا بشط فرات" — he died of thirst at the edge of the Euphrates … … I've double-checked each word, paying close attention to diacritics and potential variations in spelling. The final transcription is:
Its final transcription:
افاطم لو خلت الحسين مجدلا وقد مات عطشانا بشط فرات إذا للطمت الخد فاطم عنده واجربت دمع العين بالوجنات افاطم قومي يا ابنة الخير واندبى بجوم سموات بارض فلات
Three single-dot or spelling-variant slips on 140 characters (CER 0.02). The previous generation, amad-vlm5, read the same panel as isolated strings, never identified the poem, and invented two words on the last line ("يجوء … فلان").
What changed from amad-vlm5
- Trained with reinforcement learning on a verifiable reward. The model generates candidate transcriptions, each is scored against ground truth by character accuracy, and the behaviors that raised accuracy are reinforced.
- Failure modes gone. Repetition loops and cut-off outputs — the classic large-page failures — are what the reward punished hardest. On the historical-Arabic set where amad-vlm5 looped, CER fell from 1.45 to 0.13; on multi-line handwriting from 0.71 to 0.29.
- Better page reading. Book pages, poetry and historical prints improved (adab 0.17 → 0.11, hindawi 0.24 → 0.18, historical books 0.21 → 0.10) with the reading itself intact elsewhere.
- Sharper thinking. Reasoning got shorter (287 → 133 tokens on average) while accuracy rose: the model learned when checking pays, not just how to write more.
Fine print
Single run, single benchmark. One known weakness remains: on dense multi-line handwriting it sometimes stops after the first lines. KITAB-Bench measures transcription of images at benchmark resolution; it does not measure layout, tables as structure, or speed.
Access
amad-vlm7 is closed-weights. The model is available for evaluation and deployment through amad-iq; the open predecessors amad-vlm5 and amad-vlm6 remain public under Apache-2.0.
Murtadha · amad-iq · September 2026