LightOn

Retrieval is solved. One API to feed any model, power any agent. Three endpoints, zero config: /parse /extract /search LightOn Console dropping soon. Sandbox access, ship as you sign up! 🔗 https://lnkd.in/eMM22JM7

The data you need to benchmark enterprise RAG is the data no one shares. So we built it. Today we're releasing EDiTh on Hugging Face, an open benchmark for enterprise retrieval, designed around the questions executives actually ask. What's inside: 📄 1,004 PDFs across 6 languages and 3 formats :🎯: 36 use cases with full answer keys 🎭 5.5% plausible distractors (the kind that fool real systems) 🏢 All built around Véracier Industries, a fictional €1.8B industrial group Each use case mirrors a real executive question, the kind that usually takes a team, a week, and a stack of PDFs to answer. If you're building, buying, or evaluating RAG for the enterprise, EDiTh gives you something the public benchmarks don't: realistic documents, realistic noise, realistic stakes. Built by Adèle Guignochau and Igor Carron at LightOn. Read the release 👉🏻 https://lnkd.in/e4jrfjQ8 Dataset on Hugging Face 👉🏻 https://lnkd.in/eS_sUKh4

The plumbing era of RAG is over. Parsing. Chunking. Embeddings. Reranking. Search. Five moving parts, one fragile seam between each, and a roadmap quietly disappearing into the gaps. The teams shipping fastest have moved up the stack: 🎯 Domain expertise : the part a competitor cannot copy ⚡️ Product velocity : features instead of chunk-size debugging Retrieval infrastructure is solved. The moat is what you build on top of it. Full post → https://lnkd.in/eyUhATas More from us on this, very soon.

Radical transparency is the ultimate benchmark. It's easy to look good on a leaderboard, if you're training your model on the same data as you evaluate it on. Open source should also mean transparency. So we tested LateOn on a decontaminated BEIR, removing every document present in training data. First ColBERT above 57 NDCG@10 on standard BEIR. Still SOTA on the decontaminated version. This is the worst-case eval, yet LateOn holds. Generalization on unseen data is the only confidence that holds in production. LateOn is the retrieval core of the LightOn API, alongside LightOnOCR-2 and our multi-vector database NextPlaid. Everything is open source: check the weights, the evals, the decontamination methodology (links in comment). 🔗 Then test the full stack directly via the LightOn API: https://lighton.ai/api  

One retriever for real world data. SOTA on BEIR, built for production. LightOn open-sources LateOn, its multi-vector retriever. Apache 2.0. 149M parameters. LateOn powers LightOn's own search pipeline. Built and deployed in the best hybrid search stack on the market. Why is LateOn the retriever you actually need? We decontaminated BEIR, to remove training overlap, keep only the unseen. Late interaction compounds: LateOn tops that leaderboard too. This is the closest eval to your private data. 🔗 Read the full story: https://lnkd.in/enzQwAmH 🔑 And claim your API key to test it now: https://lnkd.in/e-S-uQgm 👨🍳 Kudos to Raphael Sourty, Antoine Chaffin, Orion Weller, Paulo Roberto Moura Júnior, Amélie Chatelain, Ph.D.!

Sovereignty is more than where your AI runs. Critical industries also demand : Air-gap capable. Fully auditable. Model-agnostic. LightOn will join Hewlett Packard Enterprise at the upcoming Defense Innovation Days 2026 in Brussels 🇧🇪 A key gathering of European defense, NATO, and industry leaders focused on one priority: building sovereign, secure, and resilient technologies for critical environments. 🔐 Sovereign AI ⚙️ Mission-ready infrastructures 🛡️ Trusted systems for defense and critical industries 📍 April 21–22, 2026 👉🏻 More details: https://lnkd.in/ex6uXqpB

From Repo to real, the gap has a name: Production. Tomorrow, LightOn joins Dell Technologies, NVIDIA, and TD SYNNEX for a full day built around one thing: getting AI to scale in the real world. ⚙️ Industrialize your use cases  🏭 Step inside an AI Factory 🤝 Activate the right partners to ship faster Not another panel. A concrete step toward deployment. 📅 Tomorrow. April 16. 9AM–5PM. 👀 Planning to attend ? come find us Or reach out to us directly 👉🏻 https://lnkd.in/eZuXsnfJ

Code search just leveled up. ColGREP 1.2.0 is out. ColGREP now combines hierarchical multi-vector search with BM25 keyword matching, so queries like "retry logic" can find both the concept and the exact function. ⚡ Faster indexing with CUDA   ⚡ Faster queries on large repositories Do more with less. 📉 35% fewer tokens consumed   🎯 ColGREP delivers more relevant results than Claude Code alone in 70% of sessions Built in the open and improved with the help of 15+ contributors. 👉 Discover our search API: https://lighton.ai/api 💻 Explore the repo: https://lnkd.in/eiguSkYV  

💫 Introducing New SOTA Long Context VLM LightOn OriOn-Qwen-SR1 reasons over full documents and executes it implicitly at inference. Reasoning is compressed into the model's weights, no verbose output, no added latency. 🥇 SOTA on MMLongBenchDoc, ahead of Qwen3 VL with 7× fewer parameters. 🙌 Kudos to Austin Veselka for this new milestone! Reasoning starts at reading. 👉 https://lnkd.in/emfzREBT

AI search & sovereign infrastructure is not a European conversation. LightOn is at GITEX AFRICA in Marrakech this week, with our partner Oreus AI. Same questions, different timezone: How do you build search that works on real enterprise data, without sending it somewhere you don't control. 🤝 If you're in Marrakech come meet David Amara, Ella Trounce, Julia Sartre and Safouane Mouraji at booth 9C-24, Hall 19. 🚀 Or start shipping now: https://lighton.ai/api