LightOn said its 150M-parameter Reason-ModernColBERT outperformed all models up to 7B parameters on BRIGHT, May 22, 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.

LightOn said on May 22, 2025 that it had released Reason-ModernColBERT, an open-source multi-vector model built for what it calls Deep Research applications. It follows GTE-ModernColBERT, the model that the release says showed state-of-the-art results on long-context retrieval. The company argued that, until then, information retrieval systems had lagged behind the boom in reasoning among large language models, lacking the reasoning capabilities needed to fully support it, and said that gap had been bridged.1^,2^,3^,4

The release said the model had 150M parameters, over 45 times smaller than certain competitors, and outperformed all models up to 7B parameters on BRIGHT, which it called the gold-standard benchmark for reasoning-intensive retrieval. It also said the model outperformed ReasonIR-8B by over 2.5 NDCG@10, a score of how well the top 10 results are ranked, on Stack Exchange queries. LightOn said the model was trained with its PyLate library in less than two hours, with fewer than 100 lines of code.5^,6^,7

Among the benefits it listed was what it called drastically reduced inference latency relative to large language models. LightOn said it was making the model, its training code and the relevant datasets publicly available, and that the model could be used on Hugging Face through PyLate.8^,9^,10

LightOn describes itself as founded in 2016 in Paris and as the first European generative AI company listed on Euronext Growth.11

The text we hold is the release in English, with its charts and pictures left out; it does not give the scores behind the BRIGHT comparison. The results are LightOn's own statements.12^,13^,5