# TII — Falcon OCR Arabic: 270M Parameters State-of-the-Art Arabic OCR

- Company: TII (tii.ae)
- Announced: 2026-10-06T06:00:00+00:00
- Category: model-update
- Coverage: not counted
- Announcement: yes
- Group: models
- Source: https://falcon-lm.github.io/blog/falcon-ocr/
- Record: https://forck.live/items/18409-falcon-ocr-arabic-270m-parameters-state-of-the-art-arabic-ocr
- Subject: Falcon LLM
- Models affected: Falcon-OCR-Arabic, Falcon OCR, Gemini 3.5 Flash, Claude Opus 5.5, Claude Fable 5, GPT Astra, GPT Sol, GPT Luna, Qwen 3.8 Max, Qwen 3.5 397B A17B, Qwen 3.5 122B A10B, Qwen 3.8 27B, Chandra OCR 2, Surya OCR 2, PaddleOCR VL 1.6, Ovis OCR 2, DeepSeek OCR, GLM OCR

TII adapted its 270M-parameter Falcon OCR model for Arabic documents through supervised finetuning and reinforcement learning. On a new 11,974-sample Arabic document benchmark, Falcon-OCR-Arabic achieved 81.9% text accuracy, second only to Gemini 3.5 Flash (84.3%), and the highest Table TEDS score (59.95%) among 17 compared models. The adaptation outperformed both larger general-purpose vision-language models and other dedicated OCR models.

## Evidence

Verbatim from https://falcon-lm.github.io/blog/falcon-ocr/:

> Falcon-OCR-Arabic scores 81.9%, second only to Gemini 3.5 Flash at 84.3%. That is a 2.5-point gap to the top model, and a lead over everything else.

---

Record: https://forck.live/items/18409-falcon-ocr-arabic-270m-parameters-state-of-the-art-arabic-ocr
Catalogue: https://forck.live/llms.txt
Current issue: https://forck.live/feed.md