TII releases 270M-parameter Falcon OCR Arabic, using SFT and RL for Arabic documents and reporting 81.87% text accuracy on its benchmark

What changed

Technology Innovation Institute introduced Falcon-OCR-Arabic on October 6, adapting its 270M-parameter early-fusion Falcon OCR architecture to Arabic in two stages: supervised fine-tuning on real and synthetic documents followed by reinforcement learning on curated high-quality samples. TII built an internal benchmark of 11,974 real-world samples across 15 document categories and reports 81.87% text accuracy and 59.95% Table TEDS, placing second in overall text accuracy behind Gemini 3.5 Flash among 17 compared models. The team also documents weaker performance on dense newspapers, magazines and comics, plus a handwriting gap. These rankings are vendor results on TII's own benchmark rather than an independent universal leaderboard.