
Independent company · Paris, France
LightOn
Paris-based, publicly listed AI company that open-sources retrieval models, ModernBERT variants and multi-vector search tools.
Paris, France
- Founded
- 2016
- Type
- Independent company
- Based in
- Paris, France
- Key papers
- 2
- Featured projects
- 8
About
LightOn was founded in Paris in 2016 by Igor Carron, Laurent Daudet, Florent Krzakala and Sylvain Gigan. Carron is chief executive and Daudet is deputy managing director. It launched an optical processing unit in 2018, and an OPU was integrated into the Jean Zay supercomputer in 2021.
The company began training language models in 2020 and says it has trained more than twelve since. It has open-sourced PAGnol for French, RITA for proteins, Mambaoutai, and models built with AnswerAI such as ModernBERT, plus retrieval models like GTE-ModernColBERT and tools such as PyLate and FastPlaid.
In November 2024 LightOn listed on Euronext Growth Paris and describes itself as Europe's first listed generative AI startup. Its commercial product is Paradigm, an on-premise enterprise platform.
Mission and ambitionsLightOn builds enterprise search and document understanding on private infrastructure while publishing retrieval models and tools as open source.
Known for
- 01
ModernBERT with AnswerAI
- 02
PyLate multi-vector retrieval
- 03
PAGnol French language models
Milestones
- 2016
Company founded
Founded in Paris by four scientists.
- 2018
Optical Processing Unit
Launched an optical hardware accelerator.
- 2020
First open LLMs
Began training and releasing language models, including PAGnol.
- 2021
Jean Zay integration
An OPU was integrated into a French national supercomputer.
- 2024
IPO and ModernBERT
Listed on Euronext Growth Paris and released ModernBERT.
- 2025
Ettin Suite
Released a family of encoder and decoder models.
Open source
Late Interaction Models Training & Retrieval (910 stars)
NextPlaid, ColGREP: Multi-vector search, from database to coding agents. (619 stars)
High-Performance Engine for Multi-Vector Search (286 stars)
RITA is a family of autoregressive protein models, developed by LightOn in collaboration with the OATML group (100 stars)
Model: 803,844 downloads, 852 likes.
Model: 16,574 downloads, 180 likes.
Dataset: 16,558 downloads, 5 likes.
Space: 129 likes on the Hub.
Research papers
Notable releases
Openness
Publishes models, datasets and libraries on Hugging Face and GitHub, mostly under MIT or Apache-2.0 for code; check model cards for model licences.
People
- Igor CarronCo-founder and CEO
- Laurent DaudetCo-founder and deputy managing director
- Florent KrzakalaCo-founder
- Sylvain GiganCo-founder
Listed from the lab's own pages and public profiles.
Contact and community
Contact address printed on the website.
Sources
Compiled from the lab's public pages, Hugging Face and GitHub. Last checked 2026-11-17. Something wrong or out of date? Tell us on Discord.



