The Technology Innovation Institute has released Falcon-OCR-Arabic, a 270M-parameter early-fusion OCR model adapted for Arabic documents. The adaptation involves supervised finetuning on real and synthetic data followed by reinforcement learning to improve accuracy on complex layouts and diacritics.
- Ranks #2 of 17 models with 81.9% text accuracy, behind only Gemini 3.5 Flash (84.3%).
- Achieves the highest Table TEDS of all compared models at 59.95%, leading the next best by 8.65 points.
- Ranks #1 on official documents, administrative forms, receipts, and invoices.
- Outperforms Claude Opus 5.5, Claude Fable 5, GPT Astra, and Qwen 3.8 Max.
The model demonstrates that adapting a compact OCR architecture can beat larger general-purpose vision-language models and other dedicated OCR systems on Arabic text.