Développer sa recherche IA en interne ou adopter une solution éprouvée ? Quels coûts, quels délais et quel retour sur investissement ?
Le mardi 24 novembre à 14h00, retrouvez nos experts pour notre webinar « IA Search souveraine : Build ou Buy ? ».
La pertinence d’une IA dépend ver la bonne information dans vos documents. Comment construire cette base tout en gardant la maîtrise de vos données ?
Au programme :
🔍 Le rôle de la recherche documentaire dans la qualité des réponses de l’IA.
⚖️ Build ou Buy : les critères pour comparer les coûts, les délais de déploiement et le retour sur investissement.
⚙️ Une démonstration live de LightOn Suite pour découvrir comment créer des workflows agentiques adaptés à vos métiers.
💬 Un temps d’échange avec nos experts pour poser vos questions et parler de vos projets.
📅 Mardi 24 novembre 2026 à 14h00 · En ligne
👉🏻 Inscrivez-vous : lnkd.in/epchT-8P
Le mardi 24 novembre à 14h00, retrouvez nos experts pour notre webinar « IA Search souveraine : Build ou Buy ? ».
La pertinence d’une IA dépend ver la bonne information dans vos documents. Comment construire cette base tout en gardant la maîtrise de vos données ?
Au programme :
🔍 Le rôle de la recherche documentaire dans la qualité des réponses de l’IA.
⚖️ Build ou Buy : les critères pour comparer les coûts, les délais de déploiement et le retour sur investissement.
⚙️ Une démonstration live de LightOn Suite pour découvrir comment créer des workflows agentiques adaptés à vos métiers.
💬 Un temps d’échange avec nos experts pour poser vos questions et parler de vos projets.
📅 Mardi 24 novembre 2026 à 14h00 · En ligne
👉🏻 Inscrivez-vous : lnkd.in/epchT-8P
0
0
5
338
The answer your team needs might be inside a chart on page 47 of a scanned report. 📄
Making it searchable starts with understanding more than the words on the page.
LightOnOCR-3 extracts data from charts and generates image descriptions, making visual information available for cument’s text.
A chart becomes structured data. An image gets a description that a search system can use.
That’s the connection between document understanding and enterprise search. LightOn Core takes the next step: indexing company information and making it retrievable for AI applications.
👉🏻 Explore LightOn Core:lighton.ai/fr/product/core
Making it searchable starts with understanding more than the words on the page.
LightOnOCR-3 extracts data from charts and generates image descriptions, making visual information available for cument’s text.
A chart becomes structured data. An image gets a description that a search system can use.
That’s the connection between document understanding and enterprise search. LightOn Core takes the next step: indexing company information and making it retrievable for AI applications.
👉🏻 Explore LightOn Core:lighton.ai/fr/product/core
0
1
3
368
What if the answer was there all along, just hidden in thousands of documents? 🐈
Oscar disappeared five years ago. But his story didn't end there.
A veterinary record under another name. A photo in an archived report. A reference in a shelter's database.
Different sources. missing cat.
With LightOn, the clues come together in under 15 minutes, with every finding linked to its source.
Because sometimes, the answer isn't missing. It's just waiting to be found.
🎬 Discover Oscar's story.
Fictional story. Real LightOn capabilities.
lighton.ai/fr
#LightOn #EnterpriseAI #EnterpriseSearch
Oscar disappeared five years ago. But his story didn't end there.
A veterinary record under another name. A photo in an archived report. A reference in a shelter's database.
Different sources. missing cat.
With LightOn, the clues come together in under 15 minutes, with every finding linked to its source.
Because sometimes, the answer isn't missing. It's just waiting to be found.
🎬 Discover Oscar's story.
Fictional story. Real LightOn capabilities.
lighton.ai/fr
#LightOn #EnterpriseAI #EnterpriseSearch
0
1
8
710
LightOn releases SlideOn & mSlideOn: open SOTA sparse retrievers at 174M and 339M params.
Dense retrieval works, but at enterprise scale it gets expensive. Every document becomes a vector, and fast vector search usually runs on GPUs.
Sparse retrieval changes the economics. Docued index, and each query only reads the documents that share its active features.
→ Search that scales to billions of documents
→ Fast search that can run on CPU
→ A smaller index to store
And quality holds up:
→ SlideOn: 56.68 on BEIR-15, above our dense model DenseOn (56.20)
→ mSlideOn vs mDenseOn: +4.76 on multilingual (MIRACL), +10.18 on long documents (MLDR), +3.25 on code (MTEB Code)
→ Prune vectors at inference to fit your latency and storage budget, with no retraining
Under the hood: we converted DenseOn and mLateOn into sparse retrievers through SAE pretraining and fine-tuning. Both models learn their own vocabulary of 32,768 concepts instead of using the tokenizer’s vocabulary.
Fully open: models, intermediate SAE checkpoints, training data and scripts.
👏🏻 Kudos to @paulomouraj and @raphaelsrty
Models: huggingface.co/collections/li…
Blog: blog.lighton.ai/blog-posts/sli…
Dense retrieval works, but at enterprise scale it gets expensive. Every document becomes a vector, and fast vector search usually runs on GPUs.
Sparse retrieval changes the economics. Docued index, and each query only reads the documents that share its active features.
→ Search that scales to billions of documents
→ Fast search that can run on CPU
→ A smaller index to store
And quality holds up:
→ SlideOn: 56.68 on BEIR-15, above our dense model DenseOn (56.20)
→ mSlideOn vs mDenseOn: +4.76 on multilingual (MIRACL), +10.18 on long documents (MLDR), +3.25 on code (MTEB Code)
→ Prune vectors at inference to fit your latency and storage budget, with no retraining
Under the hood: we converted DenseOn and mLateOn into sparse retrievers through SAE pretraining and fine-tuning. Both models learn their own vocabulary of 32,768 concepts instead of using the tokenizer’s vocabulary.
Fully open: models, intermediate SAE checkpoints, training data and scripts.
👏🏻 Kudos to @paulomouraj and @raphaelsrty
Models: huggingface.co/collections/li…
Blog: blog.lighton.ai/blog-posts/sli…
2
7
23
1,012
LightOn retweeted
We surfed the dense and late-interaction waves… now it's time to SlideOn sparse retrieval 🐕
🚀 Meet SlideOn and mSlideOn, fully open SOTA sparse retrievers at 174M and 339M parameters, trained by @LightOnIO R&D
Models:
huggingface.co/collections/li…
Blog: huggingface.co/blog/lightonai…
🚀 Meet SlideOn and mSlideOn, fully open SOTA sparse retrievers at 174M and 339M parameters, trained by @LightOnIO R&D
Models:
huggingface.co/collections/li…
Blog: huggingface.co/blog/lightonai…
7
25
74
7,495
Introducing LightOn Suite! 🚀
Enterprise AI needs more than powerful models. It needs access to the right information, and a way to turn that information into useful applications.
Today, we're launching LightOn Suite, bringing together two complementary products:
🔷 LightOn prise documents into structured, searchable knowledge that AI can use. From parsing and indexing to retrieval, with access permissions and traceability back to the original sources.
🔷 LightOn Apps : builds on Core to let teams create and deploy AI applications using their own company knowledge. From internal search and contract analysis to supplier comparisons and document workflows.
Together, they provide a shared knowledge foundation that can support multiple AI applications, without rebuilding the document infrastructure for every use case.
🤝 And we're teaming up with @Scaleway to make it available on European cloud infrastructure.
LightOn Suite is now available on @Scaleway 's French cloud, in SaaS and Managed VPC deployments. Scaleway is also making LightOnOCR-3 available through its Generative APIs, bringing our document intelligence models directly to developers.
👉🏻 Explore LightOn Core : lighton.ai/product/core
👉🏻 Discover LightOn Apps: lighton.ai/product/apps
Enterprise AI needs more than powerful models. It needs access to the right information, and a way to turn that information into useful applications.
Today, we're launching LightOn Suite, bringing together two complementary products:
🔷 LightOn prise documents into structured, searchable knowledge that AI can use. From parsing and indexing to retrieval, with access permissions and traceability back to the original sources.
🔷 LightOn Apps : builds on Core to let teams create and deploy AI applications using their own company knowledge. From internal search and contract analysis to supplier comparisons and document workflows.
Together, they provide a shared knowledge foundation that can support multiple AI applications, without rebuilding the document infrastructure for every use case.
🤝 And we're teaming up with @Scaleway to make it available on European cloud infrastructure.
LightOn Suite is now available on @Scaleway 's French cloud, in SaaS and Managed VPC deployments. Scaleway is also making LightOnOCR-3 available through its Generative APIs, bringing our document intelligence models directly to developers.
👉🏻 Explore LightOn Core : lighton.ai/product/core
👉🏻 Discover LightOn Apps: lighton.ai/product/apps
2
8
24
1,376
Got a complex document full of charts, tables, or tricky layouts? 📊
Put LightOnOCR-3 to the test!
From extracting text to understanding visual content and turning complex documents into structured, usable data, see what our latest OCR models can do.
🚀 Our demo is featured in f the Week!
Try it yourself 👇
🔗huggingface.co/spaces/lighton…
Put LightOnOCR-3 to the test!
From extracting text to understanding visual content and turning complex documents into structured, usable data, see what our latest OCR models can do.
🚀 Our demo is featured in f the Week!
Try it yourself 👇
🔗huggingface.co/spaces/lighton…
0
12
47
3,084
🔎 Is one vector enough to capture everything that matters in search?
Traditional single-vector search compresses information into a single embedding.
But what happens to the context that's lost along the way?
On October 19, our @AmelieTabatta , Head of Training & Infejoin @treygrainger and @TurnbullDoug for a technical session on Late Interaction in 2026, hosted on Maven.
They'll explore:
🔹 Why token-level embeddings can improve retrieval quality
🔹 Where late interaction makes the biggest difference 🔹 How approaches like NextPlaid are making retrieval more efficient
🔹 What it takes to scale late interaction in production
📅 October 19, 2026 | 5 PM CEST
💻 Online | Free registration
👉 Join the discussion: maven.com/p/2f6bf9/beyon…
Traditional single-vector search compresses information into a single embedding.
But what happens to the context that's lost along the way?
On October 19, our @AmelieTabatta , Head of Training & Infejoin @treygrainger and @TurnbullDoug for a technical session on Late Interaction in 2026, hosted on Maven.
They'll explore:
🔹 Why token-level embeddings can improve retrieval quality
🔹 Where late interaction makes the biggest difference 🔹 How approaches like NextPlaid are making retrieval more efficient
🔹 What it takes to scale late interaction in production
📅 October 19, 2026 | 5 PM CEST
💻 Online | Free registration
👉 Join the discussion: maven.com/p/2f6bf9/beyon…
1
7
12
1,234
World-leading document recognition. Beyond text extraction.
Introducing LightOnOCR-3, our new family of state-of-the-art AI models for document intelligence.
Building on LightOnOCR-2, it recognizes text and handwriting, describes images, extracts data from charts and preserves in a single processing step.
🏆 Leading benchmark results : ahead of Chandra-OCR-2, Mistral OCR 4.1 and dots.mocr on OlmOCR-Bench, and first among open-weight models on ParseBench.
📊 More information from every document : turn text, tables, images and charts into content your business applications can use.
⚡Compact and efficient: 0.8B and 4B versions, with document processing up to twice as fast as LightOnOCR-2.
🔓Apache 2.0 license: use, adapt and deploy the models on your own infrastructure.
From handwritten administrative files to financial reports and technical documents, make more of your information searchable and usable.
👉 Explore the models and benchmark results: huggingface.co/blog/lightonai…
👉🏻Explore the model page on our Website : lighton.ai/research/light…
#LightOn #DocumentIntelligence #OCR
Introducing LightOnOCR-3, our new family of state-of-the-art AI models for document intelligence.
Building on LightOnOCR-2, it recognizes text and handwriting, describes images, extracts data from charts and preserves in a single processing step.
🏆 Leading benchmark results : ahead of Chandra-OCR-2, Mistral OCR 4.1 and dots.mocr on OlmOCR-Bench, and first among open-weight models on ParseBench.
📊 More information from every document : turn text, tables, images and charts into content your business applications can use.
⚡Compact and efficient: 0.8B and 4B versions, with document processing up to twice as fast as LightOnOCR-2.
🔓Apache 2.0 license: use, adapt and deploy the models on your own infrastructure.
From handwritten administrative files to financial reports and technical documents, make more of your information searchable and usable.
👉 Explore the models and benchmark results: huggingface.co/blog/lightonai…
👉🏻Explore the model page on our Website : lighton.ai/research/light…
#LightOn #DocumentIntelligence #OCR
28
87
664
61,580
Ta-da ✨ The new LightOn website is live!
We’re proud to introduce a website that puts what makes AI meaningful front and center : its business applications.
Unlock the value of your documents, automate processes, turn data into actionable insights: discover how AI can real business challenges.
With this new website, we want to make these possibilities more accessible and showcase the potential of our technologies through the needs of the teams using them.
You don’t need to understand the inner workings of AI to see what it can do for your business.
👉 Explore the new website: lighton.ai
We’re proud to introduce a website that puts what makes AI meaningful front and center : its business applications.
Unlock the value of your documents, automate processes, turn data into actionable insights: discover how AI can real business challenges.
With this new website, we want to make these possibilities more accessible and showcase the potential of our technologies through the needs of the teams using them.
You don’t need to understand the inner workings of AI to see what it can do for your business.
👉 Explore the new website: lighton.ai
0
4
9
1,048
LightOn × @infocom94 : Rencontres du numérique territorial.
Pour cette première édition, @infocom94 a réuni plus d’une centaine de participants, parmi lesquels des acteurs technologiques et des représentants des collectivités, pour partager leurs problématiques, échanger sur aliser leurs connaissances et outils technologiques.
En tant que partenaire d’Infocom 94, notre Head of Customer Services & Solutions, Gauthier Zuppinger, a pris la parole lors de la table ronde sur la souveraineté numérique pour aborder ce que signifie concrètement disposer d’une technologie souveraine, ainsi que les risques et les enjeux liés à l’utilisation de technologies non souveraines.
L’ambition d’Infocom 94 d’apporter aux collectivités des solutions réalistes, pragmatiques et frugales rejoint pleinement nos valeurs et l’ADN de LightOn.
Merci à Vincent Kerbiquet, à @AmbroiseToin président d’Infocom 94, et à TerikSene pour leur confiance et leur invitation à cette première édition !
Pour cette première édition, @infocom94 a réuni plus d’une centaine de participants, parmi lesquels des acteurs technologiques et des représentants des collectivités, pour partager leurs problématiques, échanger sur aliser leurs connaissances et outils technologiques.
En tant que partenaire d’Infocom 94, notre Head of Customer Services & Solutions, Gauthier Zuppinger, a pris la parole lors de la table ronde sur la souveraineté numérique pour aborder ce que signifie concrètement disposer d’une technologie souveraine, ainsi que les risques et les enjeux liés à l’utilisation de technologies non souveraines.
L’ambition d’Infocom 94 d’apporter aux collectivités des solutions réalistes, pragmatiques et frugales rejoint pleinement nos valeurs et l’ADN de LightOn.
Merci à Vincent Kerbiquet, à @AmbroiseToin président d’Infocom 94, et à TerikSene pour leur confiance et leur invitation à cette première édition !
0
1
3
580
Building AI search is one thing. Maintaining it in production is another.
🔍 New formats, permissions, retrieval quality, infrastructure, maintenance... and what happens when the people who built it leave?
🧠 Our latest article looks at the real cost of building enterprise AI en it makes more sense to buy.
With LightOn Console, the search infrastructure is already there. Your team can focus on the application.
👉🏻 Read the article: lighton.ai/lighton-blogs/…
👉🏻 Explore Console: console.lighton.ai
🔍 New formats, permissions, retrieval quality, infrastructure, maintenance... and what happens when the people who built it leave?
🧠 Our latest article looks at the real cost of building enterprise AI en it makes more sense to buy.
With LightOn Console, the search infrastructure is already there. Your team can focus on the application.
👉🏻 Read the article: lighton.ai/lighton-blogs/…
👉🏻 Explore Console: console.lighton.ai
0
1
3
525
Une application IA, construite étape par étape et en direct.
Le 29 septembre à 11h, @LightOnIO × @ScalingoHQ vous proposent une démonstration technique basée sur des documents médicaux fictifs mais réalistes.
Au programme :
🔍 Extraire et structurer les données
🧠 Retrouver nformations utiles
📄 Générer automatiquement un dossier de sortie
☁ Déployer l’application et ses bases de données dans un environnement adapté aux données sensibles
Une démonstration concrète pour comprendre comment assembler chaque brique et passer du cas d’usage à l’application déployée.
🎥 Démo live suivie d’une session de questions-réponses
Inscription gratuite : app.livestorm.co/scalingo/de-lo…
Le 29 septembre à 11h, @LightOnIO × @ScalingoHQ vous proposent une démonstration technique basée sur des documents médicaux fictifs mais réalistes.
Au programme :
🔍 Extraire et structurer les données
🧠 Retrouver nformations utiles
📄 Générer automatiquement un dossier de sortie
☁ Déployer l’application et ses bases de données dans un environnement adapté aux données sensibles
Une démonstration concrète pour comprendre comment assembler chaque brique et passer du cas d’usage à l’application déployée.
🎥 Démo live suivie d’une session de questions-réponses
Inscription gratuite : app.livestorm.co/scalingo/de-lo…
1
3
5
629
🤝 @eurodecision × @LightOnIO : allier science de la décision et IA générative.
LightOn s’associe à @eurodecision , expert en Mathématiques Décisionnelles, Recherche Opérationnelle et IA, pour aider les organisations à mieux exploiter leurs données et leurs connaissances et à écisions grâce à l’IA.
🎯 L’objectif ? Combiner l’expertise d’EURODECISION en modélisation et optimisation avec celle de LightOn en IA générative pour développer des outils d’aide à la décision plus performants, plus rapides à déployer et plus simples à utiliser.
Cette collaboration s’inscrit notamment dans la dynamique du PIIEC-IA, qui vise à accélérer l’intégration de l’IA dans l’industrie et à faire émerger une offre européenne d’IA à forte valeur ajoutée.
Concrètement, plusieurs cas d’usage sont au cœur de cette collaboration :
- Gestion des connaissances : analyser et exploiter des documents techniques, spécifications, normes, rapports de maintenance ou données métier.
- Ingénierie logicielle : analyser les exigences et accompagner la rédaction de documentation technique et la génération de tests.
- Modélisation & optimisation : identifier les modèles et approches adaptés à des problématiques de planification et d’optimisation sous contraintes.
- Analyse & aide à la décision : expliquer les résultats des modèles en langage naturel et comparer différents scénarios « What-If ».
Un partenariat pour transformer des problématiques complexes en décisions concrètes.
👉🏻 eurodecision.com
👉🏻 lighton.ai
#IA #IAGenerative #AideALaDecision #RechercheOperationnelle #Optimisation
LightOn s’associe à @eurodecision , expert en Mathématiques Décisionnelles, Recherche Opérationnelle et IA, pour aider les organisations à mieux exploiter leurs données et leurs connaissances et à écisions grâce à l’IA.
🎯 L’objectif ? Combiner l’expertise d’EURODECISION en modélisation et optimisation avec celle de LightOn en IA générative pour développer des outils d’aide à la décision plus performants, plus rapides à déployer et plus simples à utiliser.
Cette collaboration s’inscrit notamment dans la dynamique du PIIEC-IA, qui vise à accélérer l’intégration de l’IA dans l’industrie et à faire émerger une offre européenne d’IA à forte valeur ajoutée.
Concrètement, plusieurs cas d’usage sont au cœur de cette collaboration :
- Gestion des connaissances : analyser et exploiter des documents techniques, spécifications, normes, rapports de maintenance ou données métier.
- Ingénierie logicielle : analyser les exigences et accompagner la rédaction de documentation technique et la génération de tests.
- Modélisation & optimisation : identifier les modèles et approches adaptés à des problématiques de planification et d’optimisation sous contraintes.
- Analyse & aide à la décision : expliquer les résultats des modèles en langage naturel et comparer différents scénarios « What-If ».
Un partenariat pour transformer des problématiques complexes en décisions concrètes.
👉🏻 eurodecision.com
👉🏻 lighton.ai
#IA #IAGenerative #AideALaDecision #RechercheOperationnelle #Optimisation
0
1
5
722
Your documents are not general-purpose,
Your retriever should be good beyond these
@tomaarsen recently fine-tuned mLateOn-unsupervised for medical retrieval using MIRIAD, a benchmark with 1,000 queries and 200,000 passages.
Out of the box, mLateOn reached 0.8520 NDCG@10.
After on a single RTX 3090, mLateOn-medical reached 0.9139, outperforming every general-purpose retriever tested. Qwen3-Embedding-4B, the strongest dense model in the evaluation, reached only 0.7817, despite having roughly 33 times more active parameters.
The difference comes down to how information is represented. Dense models compress an entire document into one vector. Late-interaction models preserve token-level representations, allowing them to capture signals that single-vector models may average away.
This matters even more for long, specialized documents. In this evaluation, passages averaged 941 tokens.
The conclusion is simple: strong retrieval is not only about building larger models. It is about choosing an architecture that can understand your documents and adapt to your domain.
→ Explore Tom Aarsen’s complete fine-tuning guide :
huggingface.co/blog/train-mul…
→ Put advanced retrieval to work with LightOn Console
console.lighton.ai
Your retriever should be good beyond these
@tomaarsen recently fine-tuned mLateOn-unsupervised for medical retrieval using MIRIAD, a benchmark with 1,000 queries and 200,000 passages.
Out of the box, mLateOn reached 0.8520 NDCG@10.
After on a single RTX 3090, mLateOn-medical reached 0.9139, outperforming every general-purpose retriever tested. Qwen3-Embedding-4B, the strongest dense model in the evaluation, reached only 0.7817, despite having roughly 33 times more active parameters.
The difference comes down to how information is represented. Dense models compress an entire document into one vector. Late-interaction models preserve token-level representations, allowing them to capture signals that single-vector models may average away.
This matters even more for long, specialized documents. In this evaluation, passages averaged 941 tokens.
The conclusion is simple: strong retrieval is not only about building larger models. It is about choosing an architecture that can understand your documents and adapt to your domain.
→ Explore Tom Aarsen’s complete fine-tuning guide :
huggingface.co/blog/train-mul…
→ Put advanced retrieval to work with LightOn Console
console.lighton.ai
0
4
17
996
All your agent needs is search
It gives them access to the right enterprise knowledge, grounds their responses in reliable evidence, and reduces hallucinations.
But performance alone is not enough.
For AI agents to operate at scale, they also need to make economic sense.
An is not an efficiency strategy.
By retrieving only what matters, good retrieval reduces unnecessary context, lowers token usage, and makes every inference more cost-effective. Powerful agents are useful.
Powerful and efficient agents are production-ready.
👉🏻 Start building with LightOn Console: console.lighton.ai
It gives them access to the right enterprise knowledge, grounds their responses in reliable evidence, and reduces hallucinations.
But performance alone is not enough.
For AI agents to operate at scale, they also need to make economic sense.
An is not an efficiency strategy.
By retrieving only what matters, good retrieval reduces unnecessary context, lowers token usage, and makes every inference more cost-effective. Powerful agents are useful.
Powerful and efficient agents are production-ready.
👉🏻 Start building with LightOn Console: console.lighton.ai
0
1
7
496
J-7 before the first Destination AI Coffee Chat!
On August 27, LightOn joins @TDSYNNEX and @HPE for a practical conversation on turning AI ambitions into scalable enterprise solutions.
📅 August 27, 2026
🕦 13:00 to 14:00
Register here: connect.tdsynnex.be/event/destinat…
gnAI #DestinationAI
On August 27, LightOn joins @TDSYNNEX and @HPE for a practical conversation on turning AI ambitions into scalable enterprise solutions.
📅 August 27, 2026
🕦 13:00 to 14:00
Register here: connect.tdsynnex.be/event/destinat…
gnAI #DestinationAI
0
0
2
711