UAE TII releases Falcon-OCR-Arabic: 270M parameters take second in Arabic OCR accuracy

Released by the Falcon LLM team at the UAE Technology Innovation Institute (TII) on October 6, 2026, Falcon-OCR-Arabic is an OCR (Optical Character Recognition) model purpose-built for Arabic-language documents. Without changing the original architecture, the team adapted the previously released 270M-parameter early-fusion OCR model Falcon OCR to the Arabic scenario. Its technical core follows the Falcon Perception early-fusion design: a single dense Transformer reads both image patches and text tokens in the same parameter space from the first layer, with no separate vision encoder. A hybrid attention mask implemented with PyTorch FlexAttention controls information flow: image tokens attend bidirectionally, while text tokens autoregressively decode under causal attention having observed the entire image.

The article details the unique difficulties of Arabic OCR: letters change form by position in a word (isolated, initial, medial, final); many letters share the same body shape, distinguished only by dots (ب ت ث ن ي), with a single dot potentially changing a letter or even a word; optional tashkeel diacritics are easily missed or fabricated; real documents mix right-to-left tables, Arabic with Latin letters, and Western digits (123) with Eastern Arabic digits (١٢٣); the model must also output text in logical reading order rather than visual position.

To address these challenges, the team adopted a three-stage adaptation scheme: first, build a training set with real and synthetic Arabic documents covering common categories like receipts, invoices, administrative forms, and books; then supervised finetuning (SFT) on real and synthetic data, with the next-token prediction objective making the same weights learn both to "see" Arabic glyphs and to "write" Arabic text; finally, reinforcement learning (RL) on selected high-quality samples targeting errors that token-level loss cannot easily penalize but users are highly sensitive to: dot position misalignments, fabricated diacritics, line skipping, table row duplication.

For evaluation, the team built a benchmark with 11,974 real samples covering 15 document categories, and plugged it into the OmniDocBench framework for evaluation consistency. Text accuracy is defined as 100 minus the page average normalized edit distance, and is sensitive to tashkeel, with missing or fabricated diacritics both incurring penalties.

Falcon-OCR-Arabic ranks second among 17 comparison models in text accuracy at 81.87%, behind only Gemini 3.5 Flash (84.34%), ahead of Claude Opus 5.5, Claude Fable 5, GPT Astra, Qwen 3.8 Max and all other models, and ahead of the best dedicated OCR model Chandra OCR 2 (61.67%) by over 20 percentage points. Even more striking is its table capability: Table TEDS reaches 59.95%, the highest of the 17 models and 8.65 percentage points above the second-place Gemini 3.5 Flash (51.30%). By document category, it ranks first in official documents, administrative forms, receipts, and invoices, beating Gemini 3.5 Flash; in books (90.1% vs 91.0%) and graphics (80.2% vs 82.4%) it is a close second; it is in the top four in all 15 categories. Notably, Qwen 3.5 397B A17B, with about 1,500x more parameters, scores 14.5 percentage points lower, reflecting the "small and focused" route's advantage. Comparing base to adapted: text accuracy improved from 55.39% to 81.87%, a 26.48-point gain; Table TEDS from 24.83% to 59.95%, a 35.12-point gain, nearly tripling.

The article candidly notes its limitations: on visually dense and complex layouts like newspapers, comics, magazines, and on handwritten text, it still lags Gemini 3.5 Flash. The model is available for online trial at https://ocr.arabic.aidrc.tii.ae/, with the citation format being the TII Falcon LLM team's 2026 work on 270M-parameter Arabic early-fusion OCR.