LightOn announced LightOnOCR-2, a document intelligence model with one billion parameters, on Jan 19, 2026

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, a Paris company that develops an enterprise AI platform, announced on Jan 19, 2026 a model called LightOnOCR-2, which it said it integrated into Paradigm. The company said the model could uncover and structure the information buried in enterprise documents, including the most complex ones, and called it the world’s best model for document intelligence. LightOn filed the release with the AMF, France’s financial markets regulator, under the label “inside information”, which exists for news that may move a listed company’s share price.1^,2^,3^,4^,5^,6

LightOn said the model had one billion parameters and outperformed every competing model on the OlmOCR benchmark, including those nine times its size. It said the model was compact enough to be deployed on-premise, where the documents were, and that its end-to-end architecture replaced traditional pipelines built on optical character recognition (OCR), the software that reads text from scanned pages. The company said the most critical enterprise documents, such as sensitive contracts, technical files and regulatory archives, had until then remained largely inaccessible to AI because they were too complex, too sensitive and too voluminous.7^,5^,8

LightOn presented the model as the first of three phases, which it called “Bleu, Blanc, Rouge”, and said they widened the gap between Paradigm and the rest of the market. The next phases, it said at the time, would focus on large-scale document search and then on comprehension models. Combined, it said, the technologies would let organizations deploy documentary AI integrated into their systems and fully sovereign.9^,10^,11

Igor Carron, quoted in the release, said many European companies hold a vast store of information that often goes untapped for lack of a solution suited to complex or sensitive data. He said LightOn’s ambition was to let them use it on their own infrastructure without trading performance against sovereignty, and called LightOnOCR-2 only the first building block of a broader effort to build the documentary AI infrastructure of tomorrow.12^,13^,14

The text of the release we hold gives no benchmark score and names none of the competing models. The release refers readers to a technical article on the method, benchmark, models and resources, and we hold none of the documents it refers to. The AMF’s listing for the same day also carries a record titled “Calendrier financier 2026”, French for a 2026 financial calendar; this article rests on the English release alone.15^,16^,17