#

Key takeaways

• Falcon OCR Arabic extends Falcon OCR to Arabic documents without changing the model architecture.

• The model uses supervised finetuning, then reinforcement learning on curated samples.

• It keeps an early-fusion design with one dense Transformer for image patches and text tokens.

• The output format is controlled by a prompt and a category argument.

• The training mix combines real and synthetic Arabic documents.

Falcon OCR Arabic in brief

Falcon OCR Arabic extends Falcon OCR to Arabic documents without changing its architecture. The base model is a 270M-parameter early-fusion OCR model. It was introduced in the Falcon Perception blog post.

The adaptation follows two stages. First comes supervised finetuning on real and synthetic Arabic documents. Then comes reinforcement learning on a curated set of high-quality samples. The source describes Arabic as a hard script for OCR. It also says the difficulty goes beyond the letters themselves.

How the model works

Falcon OCR reuses the early-fusion design from Falcon Perception. A single dense Transformer reads image patches and text tokens in one shared parameter space. It starts from the first layer. There is no separate vision encoder.

A hybrid attention mask controls how tokens interact. It is implemented with PyTorch FlexAttention. Image tokens attend to each other bidirectionally. Text tokens decode causally, while conditioning on the whole image.

The task is set through the prompt, not through extra modules. A category argument selects the output type. The source lists plain text, LaTeX for formulas, and HTML for tables.

Training approach

The model was trained from scratch for OCR rather than distilled from vision teachers. The source says this helps the model capture fine glyph and stroke detail. That matters for a script where a single dot can change the letter.

The training mixture includes real and synthetic Arabic documents. The source says the covered categories include receipts, invoices, administrative forms, books, and more. Real documents provide authentic layouts and real-world noise. Synthetic documents add scale, exact labels, and coverage of rare cases.

Those rare cases include dense diacritics, right-to-left tables, and mixed-script lines. This mix suggests a practical training strategy. It combines realism with controlled coverage. That is an assumption about the training design, based on the source description.

Reported results

The source gives English document scores for the model. It reports 80.3 on olmOCR and 88.64 on OmniDocBench. These numbers are presented as evidence that the model performs well beyond Arabic-only content.

The source does not provide Arabic benchmark scores in the supplied material. It also does not compare Falcon OCR Arabic with other Arabic OCR systems. So any broader performance claim would be an assumption, not a stated fact.

Why the design matters

The architecture keeps OCR focused in one shared model path. That can simplify the system compared with designs that split vision and text into separate parts. The prompt-based task control also keeps output handling flexible.

The source emphasizes fine visual detail. That is important for OCR tasks where small marks affect meaning. The model's training from scratch is presented as part of that strength. The source does not claim this is always better in every setting.

Morocco relevance

The source reports no Morocco-specific fact. A general lesson for readers is that OCR systems can benefit from mixed real and synthetic data when layouts and scripts are complex.

Operational considerations

The source points to a few practical considerations. First, the model depends on curated high-quality samples for reinforcement learning. Second, the training mix needs both real documents and synthetic coverage. Third, output behavior depends on the prompt and category argument.

These details matter for deployment planning. They suggest that document type, output format, and training data quality all influence results. The source does not provide implementation guidance beyond that.

Bottom line

Falcon OCR Arabic is a 270M-parameter OCR model adapted for Arabic documents. It keeps the early-fusion architecture, uses a two-stage adaptation process, and relies on a mixed real-and-synthetic training set. The source presents it as a strong OCR starting point for Arabic and related document layouts.