LightOn released LightOnOCR-3, document-intelligence models in 0.8-billion and 4-billion-parameter versions, under Apache 2.0

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.

On Oct 8, 2026, LightOn, an enterprise AI company founded in Paris in 2016, released LightOnOCR-3, a family of AI models for reading documents. The models came in 0.8-billion and 4-billion-parameter versions and were released under the Apache 2.0 license. OCR, short for optical character recognition, is software that turns text in scans and images into digital text; the company said LightOnOCR-3 goes beyond that, describing images, extracting data from charts and preserving document structure. LightOn filed the announcement with the AMF, the French financial markets regulator, as inside information, the category for news that a listed company must publish promptly.1^,2^,3^,4^,5^,6

On OlmOCR-Bench, which evaluates document text extraction, the company said LightOnOCR-3 ranked second, behind Infinity-Parser2-Pro, a model more than eight times its size, and ahead of Chandra-OCR-2, Mistral OCR 4.1 and dots.mocr. It said the model led open-weight models on ParseBench, which evaluates document layout, image descriptions and structured data extraction from charts and scientific figures, and ranked first on FRBench-pdf2md, a benchmark of French-language documents, with its largest gains in handwriting recognition. The company said the models could process documents up to twice as fast as LightOnOCR-2, the model they build on, and were the cheapest among leading OCR models. Run on an organization's own infrastructure, it said, processing costs could fall below one cent per thousand pages, depending on hardware, usage levels and workload. LightOn said LightOnOCR-2 had 240,000 monthly downloads on Hugging Face.3^,7^,8^,9^,10^,11^,12^,13

LightOn named financial analysis, administrative workflows and technical document search as uses. The license allowed commercial use, modification and redistribution, and organizations could run the models on their own infrastructure.14^,15^,16^,17

The rankings, speed and cost claims are LightOn's own, and the release points to a technical article for the detailed scores. Our records hold the text of this one release, 5,445 characters on 4 pages, in English, with its pictures, charts and layout left out, and none of the documents it refers to, so that article's scores are not in our records.18^,19