LightOn announced BioClinical ModernBERT, a medical-text encoder built on ModernBERT, on Jun 13, 2025

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 Jun 13, 2025, LightOn announced BioClinical ModernBERT, a new encoder model for medical language tasks. The company said the model outperformed all existing encoders on medical classification and named entity recognition tasks, and said it set a new state-of-the-art benchmark for medical NLP (natural language processing) applications. LightOn filed the release with France's financial markets regulator, the AMF, as inside information, the category for news that a listed company must publish promptly.1^,2^,3^,4

The release said Thomas Sounack of the Dana-Farber Cancer Institute trained the model in collaboration with researchers at LightOn, Harvard University, MIT, McGill University, Albany Medical College and Microsoft Research. The model started from ModernBERT, which the release described as a recent release by LightOn and AnswerAI, and continued its pre-training on what the company described as an extensive collection of medical texts. LightOn said a scheduling feature of ModernBERT let researchers continue pre-training without cold restarts. It added that long-context support matters because medical texts such as full clinical notes and large reports can be very long.5^,6^,7^,8^,9

LightOn also presented the work as a method others could reuse: it said the approach was reproducible, and that the model's transfer to new domains opened the door for anyone seeking to tailor ModernBERT to their own specialized data. In the release LightOn described itself as a pioneering player in sovereign generative AI and the first European generative AI company listed on Euronext Growth, and its Paradigm platform as a way for organizations to deploy AI at scale while keeping their data confidential.10^,11

The release we hold runs 3 pages in English, with its pictures and layout left out, and we hold none of the documents it refers to. The text we hold gives no benchmark scores, so the evidence behind the claim of outperforming existing encoders is not in the material we hold.12^,13