LightOn said it adapted its LightOnOCR-2 model to Arabic, in a demonstration based on 12,000 synthetic pages

Written by Claude Sonnet 5.5, an AI model, from public records. Each figure in it was checked automatically against the source's text or our database before it was published.

LightOn said on Jun 12, 2026 that it had adapted LightOnOCR-2, its document understanding model, to Arabic through fine-tuning. The company said the demonstration used a dataset of 12,000 synthetic pages with their reference transcriptions, produced with a modified version of its synthetic document generator. The AMF, France’s financial markets regulator, lists the release as inside information, the label for news that a listed company has to publish as soon as it can.1^,2^,3^,4

LightOn said Arabic poses specific challenges for OCR, optical character recognition, the reading of text from scanned documents: the script runs from right to left, characters connect in cursive form, and open datasets and specialized models are less widely available than for Latin-based languages. It said those limits can slow the automation of document workflows for organizations that process archives and administrative, legal or heritage documents in Arabic. The company said it was making guides to reproduce the fine-tuning available on its Hugging Face space, and that the work addressed needs it had met in the Middle East, where it was already working with public- and private-sector organizations.5^,6^,7^,8

LightOn said LightOnOCR-2 was released as open source under the Apache 2.0 license and played a central role in the production document ingestion process within LightOn Console, its self-service offering. It cited more than 3 million downloads of the model, along with fine-tunings already carried out by the community. The company put the base model’s score on OlmOCR-Bench at 83.2%.9^,10^,11

The AMF’s English record of that day carries a title about enhancing LightOn’s Paradigm platform with real-time web access through a strategic partnership with Linkup, and its French record a title about adapting LightOn’s OCR model to Arabic. The English text we hold repeats the Linkup headline, but its paragraphs do not mention Linkup or web access, and our records do not hold a text that announces that partnership.12^,13^,14^,15