LightOn

2.4K posts

LightOn banner
LightOn

LightOn

@LightOnIO

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

Paris, France Katılım Ekim 2015
830 Takip Edilen4.6K Takipçiler
Sabitlenmiş Tweet
LightOn
LightOn@LightOnIO·
The multi-vector era is here and there is no going back. Reason-ModernColBERT tops BrowseComp-Plus, the hardest agentic search benchmark available, by 7.59 points on accuracy. 🥇on accuracy. 🥇on recall. 🥇on calibration. 📉 Fewest search calls. The models it outperforms? Up to 54× larger. Reasoning-intensive retrieval (BRIGHT), code search (MTEB Code), agentic Deep Research (BrowseComp-Plus). The pattern is the same: late interaction dominates, with a fraction of the parameters. 149M parameters. Open weights. Open code. Built with PyLate in a few hours. Full results, analysis and recipe on LightOn blog: lighton.ai/lighton-blogs/…
LightOn tweet media
English
2
20
139
22.2K
LightOn
LightOn@LightOnIO·
“Current approaches won't fix search” At #ECIR2026, LightOn is doubling down on Encoders and Late Interaction: Open Source Search at LightOn 🔧 Co-organizing the LIR workshop 📊 Presenting ColBERT-Zero 🎤 Industrial Day talk: Encoders and Late Interaction We’ll detail our open source work on: ⚙️ The case for modern encoders (ModernBERT, Ettin) 🔍The limits of single-vector search 🚀 Late interaction with PyLate, with open models outperforming systems up to 52× largerSame goal. Different approach. Make search actually work. catchup with @AmelieTabatta & @antoine_chaffin @raphaelsrty and Paulo Roberto de Moura Júnior for a late (or dense) coffee Want to test multi-vector retrieval in practice? →lighton.ai/pricing
LightOn tweet media
English
0
4
7
561
Rasmus Toivanen
Rasmus Toivanen@RasmusToivanen·
@IgorCarron @LightOnIO While I say great job as European, and while impressive I would not be banging my chest on single benchmark, kinda niche thing. Get SaaS API (If you do not already) and tell you are outgrowing something like Azure Doc intelligence in EU then that would be great
English
1
0
1
1.8K
Igor Carron
Igor Carron@IgorCarron·
Everyone told us the AI race was over. That Europe🇪🇺 missed it. That you need $10B clusters and closed-source moats to compete. Then @LightOnIO's LightOnOCR-2 -1B parameters, open-source, running on a single GPU you can put on your desk- just beat OpenAI GPT-5 mini, Anthropic Claude Sonnet, Google Gemini 2.5 Flash, Zhipu GLM-4.5V, and DeepSeek-OCR on table extraction. The work that actually matters. Not Silicon Valley 🇺🇸 Not Shenzhen🇨🇳 Not Beijing 🇨🇳 Not Hangzhou 🇨🇳 From Paris🇫🇷 ...with love 💕 The race isn't over. It never was.
Igor Carron@IgorCarron

x.com/i/article/2037…

English
21
52
424
46.3K
LightOn
LightOn@LightOnIO·
@iledefrance × @LightOnIO 30% de tickets IT en moins -> 360k€ économisés par an. 15 000 à 20 000 tickets IT par mois. Une grande partie ne nécessite pas d’intervention technique, mais simplement un accès rapide à la bonne information Avec LightOn, déployé sur une infrastructure souveraine, la Région a lancé un assistant IA interne désormais utilisé par plus de 3 000 agents. Résultats : ⚙️30 % de réduction prévue des tickets IT 🔐 Déploiement entièrement souverain (on-premise) 🔗Intégration avec ServiceNow et SSO 📈Environ 360 k€ d’économies annuelles Un exemple concret d’une IA qui résout des frictions opérationnelles, au-delà de l’expérimentation technologique. Découvrez ce cas d'usage complet 👇🏻 lighton-dev.webflow.io/fr-blog-posts/…
LightOn tweet media
Français
0
5
12
353
LightOn
LightOn@LightOnIO·
🎙️ "Il faut penser l'IA comme une infrastructure ancrée dans la réalité documentaire des organisations, et non plus comme une application ex machina." @IgorCarron était l'invité de @simottel sur @bfmbusiness pour revenir sur les dernières innovations de LightOn et ce nouveau champ qu'elles ouvrent : l'intelligence documentaire. bfmtv.com/economie/repla…
Français
0
3
8
525
LightOn
LightOn@LightOnIO·
To everyone who has hit the wall doing RAG: we planned this one for you. Broken retrieval. Hallucinations at inference. Pipelines that fold the moment data gets sensitive. We know where it breaks. We built Paradigm to fix it. LightOn is joining @TDSYNNEX on March 27, to show what production-grade retrieval actually looks like inside a regulated enterprise: 🔍 Hybrid search, 🧠 structured reasoning, 📋 full auditability, 🔒 zero data leaving your infra. @Gauthier_Z brings the technical depth alongside Fabrice Bagniakana. No skipping the hard parts. 📅 March 27 · 11:00–12:00 CET 🔗 Register: @7fe14ab6-8f5d-4139-84bf-cd8aed0ee6b9" target="_blank" rel="nofollow noopener">events.teams.microsoft.com/event/21406722…
LightOn tweet media
English
0
3
8
504
LightOn
LightOn@LightOnIO·
LightOn bet on multi-vector early. This is pay day. When most systems were still compressing everything into a single embedding, LightOn went the other way. We built the ecosystem, open source from the ground up, and multi-vector is now winning where it counts: 🧩 Complex queries. 📚Long documents. 💻 Code. 🎯 Out-of-distribution. 🤖 Agentic systems. @AmelieTabatta and @antoine_chaffin joined @CShorten30 on the @weaviatepodcast to break down why we made this bet, what we've built, and what it unlocks for the next generation of search and reasoning. 🎧: youtu.be/44GC3E-WbHU
YouTube video
YouTube
LightOn tweet media
English
2
11
22
1.2K
LightOn
LightOn@LightOnIO·
Days since LightOn last shipped a retrieval milestone: 0 BM25x just dropped. Don't choose between lexical, dense and multi-vector semantic retrieval. Run all three. They're cheaper, faster, better simultaneously. That's hybrid search with no compromises → lighton.ai/lighton-api
Raphaël Sourty@raphaelsrty

Released BM25x on @LightOnIO git this week, 13000 queries per second (QPS) on MSMARCO (8.8M documents) with 4*H100 against 19 QPS for BM25s (CPU). The comparison is not fair but let me introduce bm25x 👇

English
1
11
76
7.1K
LightOn retweetledi
Raymond Weitekamp
Raymond Weitekamp@raw_works·
currently re-embedding my entire machine, thank you very much! LateOn-Code-edge for code search and Reason-ModernColBERT for prose/docs search.
Antoine Chaffin@antoine_chaffin

BrowseComp-Plus, perhaps the hardest popular deep research task, is now solved at nearly 90%... ... and all it took was a 150M model ✨ Thrilled to announce that Reason-ModernColBERT did it again and outperform all models (including models 54× bigger) on all metrics

English
3
8
42
5.2K
LightOn retweetledi
Connor Shorten
Connor Shorten@CShorten30·
Super exciting win for Agentic Search and Late Interaction! 🧬 GPT-5 + Reason-ModernColBERT (150M) reaches ~88% accuracy with an average of ~13 search calls. For reference, when BrowseComp-Plus was published in August 2025, the max accuracy reported was ~70% using GPT-5 + Qwen3-Embed-8B, using ~22 search calls. Searching with reasoning 🤖💭is a beast. 🔥 This is a huge evangelist for semantic search and Late Interaction models are particularly shining thanks to their effectiveness at long input modeling with fine-grained similarity scores. 🛠️ Congratulations @antoine_chaffin and team! 🎉
Antoine Chaffin@antoine_chaffin

BrowseComp-Plus, perhaps the hardest popular deep research task, is now solved at nearly 90%... ... and all it took was a 150M model ✨ Thrilled to announce that Reason-ModernColBERT did it again and outperform all models (including models 54× bigger) on all metrics

English
3
14
103
8.7K

Keşfet

@AmelieTabatta @antoine_chaffin @raphaelsrty @RasmusToivanen @IgorCarron @iledefrance @simottel @bfmbusiness