LightOn

LightOn is a French artificial-intelligence company founded in Paris in 2016 by Igor Carron, Laurent Daudet, Florent Krzakala and Sylvain Gigan. Its early work centered on optical computing: using the physical propagation and interference of light to perform operations relevant to machine learning. The company developed an Optical Processing Unit (OPU), intended to accelerate highly parallel workloads such as matrix operations while potentially reducing the energy and hardware requirements associated with conventional electronic computation. The OPU was launched in 2018 and became one of the most visible expressions of LightOn's original hardware strategy. In 2021, an OPU was integrated into Jean Zay, a high-performance computing system operated by the French National Centre for Scientific Research. This made Jean Zay the first computer listed in the TOP500 ranking of the world's most powerful supercomputers to use this type of optical-processing hardware. The deployment connected LightOn's technology with scientific computing and demonstrated an application beyond laboratory research. From around 2020, LightOn broadened its activity toward machine-learning models and enterprise software. It trained and released language models on supercomputing infrastructure, while continuing to research efficient approaches to large-scale AI. The company's current principal offering is Paradigm, a secure generative-AI platform designed for enterprise deployment. Paradigm can be operated on premises or in controlled environments, a positioning intended for organizations that need to keep sensitive information within their own infrastructure and maintain governance over models, data, access and retrieval systems. LightOn's software activity includes language models, document and information-retrieval technology, and open-source tools for retrieval-augmented generation and AI-agent applications. Products and research outputs associated with the company include ModernBERT, developed with AnswerAI; MonoQwen-Vision, described as a visual-document reranker; the Reason-ModernColBERT and GTE-ModernColBERT models; BioClinical ModernBERT, developed with academic and research partners; and the Ettin family of encoder-decoder models developed with Johns Hopkins University. The company has stated that it trained more than 12 large language models between 2020 and 2025. A further part of LightOn's technical proposition is multi-vector retrieval. Its open-source PyLate, FastPlaid and PyLate-rs tools are intended to support fast, updateable search systems for enterprise knowledge bases, recommendation systems and retrieval-augmented-generation pipelines. FastPlaid is positioned as a higher-performance implementation related to Stanford's PLAID approach, while PyLate-rs brings retrieval capabilities to browser environments through WebAssembly. LightOn's customers and implementation partners have included public-sector and industrial organizations. After the end of 2024, reported deployments and applications included the Île-de-France region, Safran, Groupama and France's space agency, CNES. The company also announced relationships with Orange Business and Hewlett Packard Enterprise to connect its software with cloud, infrastructure and enterprise-technology offerings. In November 2024, LightOn carried out an initial public offering on Euronext Growth Paris. The listing was described as making it Europe's first listed generative-AI startup. The move marked a transition from a venture-backed deep-technology company focused primarily on optical acceleration toward a public enterprise-AI provider whose commercial center of gravity is secure generative AI, retrieval and language-model software. LightOn remains active in both applied AI software and research into efficient computing architectures.

LightOn was established in Paris in 2016 by Igor Carron, Laurent Daudet, Florent Krzakala and Sylvain Gigan. The founders combined expertise in optics, physics, computer science and machine learning with the goal of developing new hardware for AI computation. Their initial proposition was based on optical processing, in which light propagation performs or assists mathematical operations that are central to neural-network workloads. This approach was intended to exploit the natural parallelism of optical systems and address the speed and energy demands of increasingly large models. The company's first major commercial technology was the Optical Processing Unit, launched in 2018. The OPU was designed as an accelerator for computationally intensive and massively parallel tasks. Rather than replacing all conventional computing infrastructure, it was conceived as a specialized component that could work alongside existing systems. LightOn's optical approach became particularly relevant as machine-learning models required larger matrix operations and more data movement. A significant validation of the hardware strategy came in 2021, when an OPU was integrated into Jean Zay, a French high-performance computing system operated by the National Centre for Scientific Research. Jean Zay became the first supercomputer appearing in the TOP500 ranking to use such optical-processing hardware. The integration demonstrated that LightOn's technology could be connected to a production-scale scientific-computing environment rather than remaining solely a research prototype. During the same broad period, LightOn began expanding beyond hardware. From 2020 onward, the company trained a series of large language models and released some models as open source using supercomputing resources. This work gave the company experience across model training, inference, retrieval and deployment. As generative AI became more commercially important, LightOn increasingly emphasized enterprise software rather than presenting optical acceleration as its only product category. The company's principal current platform is Paradigm, an enterprise generative-AI system designed for secure and controlled deployment. The platform is positioned for on-premises or otherwise protected environments where organizations need to use language models without sending sensitive information to an uncontrolled public service. Its scope includes enterprise knowledge access, model interaction, retrieval-augmented generation and related governance requirements. This orientation has helped LightOn target public institutions and regulated or information-sensitive industries. LightOn's model and retrieval research has developed alongside Paradigm. The company worked with AnswerAI on ModernBERT and released MonoQwen-Vision for visual-document reranking. In 2025 it announced additional models, including Reason-ModernColBERT, GTE-ModernColBERT and BioClinical ModernBERT. The biomedical model involved collaboration with Dana-Farber Cancer Institute and researchers affiliated with Harvard University, MIT, McGill University, Albany Medical College and Microsoft Research. LightOn also released the Ettin Suite, described as a state-of-the-art paired encoder-decoder model family developed with Johns Hopkins University. The company has also invested in open-source retrieval infrastructure. PyLate supports multi-vector model training, while FastPlaid is intended to make multi-vector retrieval practical for large, dynamic enterprise collections. PyLate-rs extends related capabilities to browser applications through WebAssembly. These tools are aimed at real-time enterprise search, recommendation systems, retrieval-augmented-generation pipelines and AI agents that must work against changing information rather than a static training corpus. LightOn's commercial activity broadened through deployments and partnerships. Reported users or implementation environments after the end of 2024 included the Île-de-France region, Safran, Groupama and CNES. The company also worked with Orange Business and Hewlett Packard Enterprise to integrate its software with cloud and enterprise infrastructure. These relationships reflect its movement toward enterprise delivery and ecosystem integration. In November 2024, LightOn completed an IPO on Euronext Growth Paris. It was characterized as Europe's first listed generative-AI startup. The listing represented a major corporate transition, giving the company access to public-market financing and visibility while it shifted from an optical-computing specialist toward a broader provider of secure generative-AI software, language models and retrieval technology. LightOn continues to retain its optical-computing heritage while its current commercial identity is centered on enterprise AI.

• 2025

  New model and retrieval releases

  [1]LightOnen.wikipedia.org

  LightOn releases several advanced retrieval and biomedical models and the open-source Ettin Suite.

• 2024

  Initial public offering on Euronext Growth Paris

  [1]LightOnen.wikipedia.org

  LightOn becomes a publicly traded company and is described as Europe's first listed generative-AI startup.

• 2021

  OPU integrated into Jean Zay

  [1]LightOnen.wikipedia.org

  An OPU is installed in Jean Zay, the first TOP500-listed supercomputer to use this type of optical hardware.

• 2020

  Expansion into language-model research

  [1]LightOnen.wikipedia.org

  LightOn begins training and releasing large language models using supercomputing infrastructure.

• 2018

  Optical Processing Unit launched

  [1]LightOnen.wikipedia.org

  LightOn introduces the OPU, an optical accelerator intended for massively parallel machine-learning computation.

• 2016

  LightOn is founded in Paris

  [1]LightOnen.wikipedia.org

  Igor Carron, Laurent Daudet, Florent Krzakala and Sylvain Gigan establish the company to develop optical-computing and AI technologies.

Secure, enterprise-focused generative-AI and retrieval technology provider with origins in optical computing and AI acceleration.

• Enterprise generative-AI platform

  Paradigm

  Paradigm is LightOn's principal current commercial platform. It provides a secure environment for enterprise generative AI, including controlled access to language models, organizational knowledge and retrieval-based applications. Its on-premises or protected-deployment orientation is aimed at customers that need stronger control over confidential data, infrastructure and model use than a general public chatbot typically provides.

• AI accelerator hardware · 2018

  Optical Processing Unit

  The OPU is LightOn's original optical-computing accelerator. It uses optical processing to address highly parallel mathematical operations relevant to machine learning. The technology was developed to improve computational efficiency and was integrated into the Jean Zay supercomputer in 2021, demonstrating its use in a high-performance scientific-computing environment.

• Language model

  ModernBERT

  ModernBERT is a language-model project released by LightOn in collaboration with AnswerAI. It belongs to the company's broader effort to develop and distribute models for enterprise search, text understanding and retrieval-oriented applications.

• Visual-document retrieval model

  MonoQwen-Vision

  MonoQwen-Vision is a visual-document reranking model associated with LightOn's multimodal and retrieval work. It is intended to improve the ordering and relevance of results involving documents whose information is conveyed through both text and visual structure.

• Retrieval models · 2025

  ModernColBERT model family

  The ModernColBERT family includes models such as Reason-ModernColBERT and GTE-ModernColBERT. These models support dense or multi-vector retrieval use cases, including enterprise search, retrieval-augmented generation and applications in which relevance must be calculated over large, changing collections.

• Biomedical language model · 2025

  BioClinical ModernBERT

  BioClinical ModernBERT is a biomedical and clinical language model developed with Dana-Farber Cancer Institute and a wider group of academic and research collaborators. It applies LightOn's language-model work to specialized medical and biomedical text rather than general-purpose enterprise content alone.

• Open-source encoder-decoder models · 2025

  Ettin Suite

  The Ettin Suite is an open-source family of paired encoder-decoder models developed with Johns Hopkins University. LightOn describes it as a state-of-the-art model suite intended to broaden access to capable language-model architectures.

• Open-source retrieval software

  PyLate

  PyLate is an open-source toolkit for training and working with multi-vector retrieval models. It is designed for enterprise search and retrieval-augmented systems in which multiple representations can improve matching between queries and complex documents.

• Multi-vector retrieval software

  FastPlaid

  FastPlaid is LightOn's high-performance implementation for multi-vector retrieval. It is intended to make retrieval suitable for production workloads such as dynamic enterprise search, recommendations and AI-agent knowledge access.

• Browser-based retrieval software

  PyLate-rs

  PyLate-rs extends LightOn's retrieval tooling to browser-based applications through WebAssembly compilation, enabling related multi-vector retrieval capabilities to run in web environments.

• 2024

  Complete an initial public offeringOther

  Background

  LightOn sought public-market access as its commercial focus moved toward enterprise generative-AI software.

  Action taken

  The company launched an IPO and listed on Euronext Growth Paris in November 2024.

  Aftermath

  LightOn became a publicly traded AI company and was described as Europe's first listed generative-AI startup.

  Source

• 2020

  Expand from optical hardware into language models and enterprise AIStrategy

  Background

  The growth of large language models created an opportunity to apply LightOn's technical capabilities to software, model training and enterprise deployment.

  Action taken

  The company began training and releasing language models and later centered its commercial offering on the secure Paradigm platform.

  Aftermath

  LightOn developed a broader product portfolio spanning generative AI, retrieval, open-source models and enterprise software.

  Source

• 2016

  Adopt optical computing as an AI-acceleration strategyStrategy

  Background

  The founders sought to address the computational and energy demands of machine learning through optical processing and hardware designed for parallel operations.

  Action taken

  LightOn developed the OPU and related optical-computing technologies as its initial product direction.

  Aftermath

  The approach led to an OPU deployment in the Jean Zay supercomputer and established the company's original deep-technology identity.

  Source

• 2026.10.13

  Reported2026.10.13

  Scaleway launches LightOnOCR-3 managed API within its Generative APIs[3]Scaleway在其生成式API中推出LightOnOCR-3托管API服务finanznachrichten.de

  On October 13, 2026, Scaleway launched its first OCR offering, a LightOnOCR-3 endpoint within its Generative APIs, becoming the first cloud provider to offer LightOn's open-source document intelligence model as a managed API, in 0.8-billion and 4-billion-parameter versions, on a European cloud.

• 2026.10.13

  Reported2026.10.13

  LightOn launches LightOn Suite platform for building AI applications on enterprise data[3]Scaleway在其生成式API中推出LightOnOCR-3托管API服务finanznachrichten.de

  On October 13, 2026, LightOn launched LightOn Suite, a platform combining LightOn Core, which makes enterprise document assets usable by AI, and LightOn Apps, which turns that knowledge into business applications. The suite is available as SaaS and as a Managed VPC hosted on Scaleway's French infrastructure.

Read more

1. 1LightOnen.wikipedia.org
2. 2LightOn official websitelighton.ai
3. 3Scaleway在其生成式API中推出LightOnOCR-3托管API服务finanznachrichten.de