LightOn

LightOn is a leading European generative AI company delivering secure on-prem RAG for document intelligence, enabling safe use of sensitive data behind firewall

[LightOn @LightOnIO]

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. Documents go into an inverted 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: https://t.co/UNK78RhTNv
Blog: https://t.co/uVVs9JX5pb

[Paulo Moura @paulomouraj]

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:
https://t.co/fedmPHd15A
Blog: https://t.co/b8oiNDSGRz

[LightOn @LightOnIO]

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 Core : transforms enterprise 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 : https://t.co/F64QagU1EZ

👉🏻 Discover LightOn Apps: https://t.co/2GZeHTZSBV

[LightOn @LightOnIO]

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 Hugging Face's Spaces of the Week!

Try it yourself 👇
🔗https://t.co/UEc2yWJGvO

[LightOn @LightOnIO]

🔎 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 & Inference at LightOn, will join @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: https://t.co/Xt9QIxk8jz

[LightOn @LightOnIO]

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 document structure, all 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: https://t.co/xXJD1Eo9XA

👉🏻Explore the model page on our Website : https://t.co/y5ZWMpMOL9

#LightOn #DocumentIntelligence #OCR

[LightOn @LightOnIO]

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 address your company’s 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: https://t.co/OxRzOnGqRv

[LightOn @LightOnIO]

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 leurs pratiques et mutualiser 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 !

[LightOn @LightOnIO]

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 search in-house, and when 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: https://t.co/3U1aKPut6I

👉🏻 Explore Console: https://t.co/RNJQKEHzQ2

[LightOn @LightOnIO]

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 et contextualiser les informations 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 : https://t.co/8BorrEaR6e

[LightOn @LightOnIO]

🤝 @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 prendre de meilleures décisions 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.

👉🏻 https://t.co/JQm9wR4rjy
👉🏻 https://t.co/F8SZUxHacy

#IA #IAGenerative #AideALaDecision #RechercheOperationnelle #Optimisation

[LightOn @LightOnIO]

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 a short fine-tuning run 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 :
https://t.co/Z1mlZ4DgVA

→ Put advanced retrieval to work with LightOn Console
https://t.co/kY55OAXWIq

[LightOn @LightOnIO]

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 infinite context window 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: https://t.co/kY55OAXWIq

[LightOn @LightOnIO]

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: https://t.co/a12cCnO6gZ

#EnterpriseAI #SovereignAI #DestinationAI

[LightOn @LightOnIO]

RAG quality is decided before the prompt reaches the LLM.

LightOn’s open-source stack strengthens every layer between raw documents and grounded answers:

📄 LightOnOCR-2 for parsing at scale
🔍 mLateOn, a state-of-the-art retriever
🎯 LightOn-rerank, a state-of-the-art reranker

Better parsing. Sharper retrieval. Cleaner context.

Use each component independently, or access the full production-ready pipeline through LightOn Console.
The bricks are already in your dependency tree.

Start building: https://t.co/Bn9F6pK6Zx

[LightOn @LightOnIO]

Your AI doesn't need more context.
It needs the right receipts, contract clauses, meeting transcripts and diagrams
📄

LightOn Search grounds every answer in your organization's truth by retrieving the most relevant evidence before it reaches your LLM.

Less noise.
Fewer tokens.
Sourced answers from the documents that actually matter. Always.

The best retrieval system doesn't find more. It finds what counts.

⚡ Start building with free credits: https://t.co/kY55OAXWIq

#RAG #Search #EnterpriseAI #Developers

[LightOn @LightOnIO]

Great search deserves a real test.

Public datasets are now available directly in LightOn Console, so you can test our search on real-world document collections from the start.

Explore the data, run semantic searches, ask questions, and assess retrieval performance without downloading, cleaning, chunking, or indexing documents first.

Less setup. More experimentation. Faster prototypes.

Start now 👉 https://t.co/H8TLnEWqYa