LightOn said its open-source GTE-ModernColBERT beat the best existing models by 10 points on long documents in LongEmbed, May 7, 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 7, 2025 that it had released GTE-ModernColBERT, an open-source multi-vector retrieval model. A retrieval model is the search step that finds relevant passages for AI-generated answers, as in retrieval-augmented generation (RAG); the release said the model enabled more accurate retrieval for such responses. The release contrasted multi-vector models with single-vector embedding models, which it said had become standard in the industry.1^,2^,3^,4

The company said the model outperformed other models in the field, among them those of Alibaba, Snowflake, Cohere, BAAI and JinaAI, on LongEmbed, a benchmark it called industry-standard. It put the lead over the best existing models at a 10 point margin on documents of up to 32,000 tokens, which the release equated to texts of dozens of pages, even though the model was first trained only on 300 token excerpts from the MS MARCO dataset. LightOn also listed as a benefit that it was the first model to surpass ColBERT-small on the BEIR benchmark.5^,6^,7

LightOn said it created the model by leveraging the ModernBERT architecture, which it said it built, and its PyLate library, a framework for training multi-vector retrieval models that it said needed only 80 lines of code to reproduce the training. It said that training on MS MARCO by knowledge distillation produced a lightweight model that does not compromise on performance. The release added that most major vector databases, among them QDrant, LanceDB, Weaviate and Vespa, supported multi-vector indexing.8^,9^,10^,11^,12

The release said the model was released under an Apache 2.0 licence and that PyLate was open source as well, and it pointed readers to Hugging Face to try the model. LightOn describes itself as founded in 2016 in Paris and as the first European generative AI company listed on Euronext Growth.13^,14^,15

The text we hold is the release in English; the AMF also lists a French record of the same day. The results are LightOn's own, and the text we hold does not state which measure the 10 point margin is expressed in.16^,17^,18^,6