IdePHICS lab

"Information Learning & Physics" is an academic lab. in EPFL, Switzerland, focusing on computer science, machine learning & statistical physics. PI: @krzakalaF

[Sylvain Peyronnet @speyronnet]

Nouvelle chaine YT pour @ibou (abonnez vous !)
Premier podcast QUEDUWEB avec comme invitée @AmelieTabatta de @LightOnIO, coanimé par @Greg0ry et moi-même.

Sortez vos plus beaux likes ! Ca parle LLM, RAG, etc.

https://t.co/2ekWWZnCaa

[LightOn @LightOnIO]

Today, we're introducing LightOn Console.

⚙️ Three endpoints:
/Parse any documents
/Extract structured data
/Search enterprise knowledge with citations

🔌 Built-in connectors. MCP-ready. Governance enforced at the chunk level.

No infrastructure. No pipeline maintenance. No dedicated retrieval team required.
Make your enterprise knowledge agent-readable now!

Read the launch announcement: https://t.co/LcxXqyOgo5

Test it now: https://t.co/RNJQKEHzQ2

[Amélie Chatelain @AmelieTabatta]

> Goes to buy @le_Parisien
> News stand man asks why this one in particular
> Me: I'm in it :D!!!

[Mary Newhauser @m_newhaus]

Some of the most exciting small models out there are by @LightOnIO.

Been following their work for years and so happy to see them start to break into the mainstream.

[LightOn @LightOnIO]

50 million downloads on @huggingface!

LightOn SOTA late-interaction and dense retrievers, OCR models, and LLMs are validated by the community and tested every day in production.

🧪 LightOn is now one of the most active labs in the world in retrieval, pushing the Pareto frontier of AI, always open source.

🙌 Huge kudos to @iacopo_poli, @AmelieTabatta, @antoine_chaffin, @raphaelsrty, @staghado, @CavaillesAdrien, @baptaubertin, Maxence Lasbordes, @paulomouraj and the entire team for the work!

🔥 As proud as LightOn is of this milestone, the model is only one building block. Orchestration is what makes AI systems work.

Now it’s your turn to test them in production: https://t.co/RNJQKEHzQ2

[Igor Carron @IgorCarron]

50,000,000 downloads @LightOnIO
LightOn's open R&D powers the full document intelligence stack from📄 OCR to🔎 Indexing to⚡ Late-interaction search to reason.
🎁Featured in one product: https://t.co/cpTz8IzxQH
🇪🇺 European AI sovereignty, shipped.

[Nicolas Bouzou @nbouzou]

Du Brexit à Trump en passant par Raoult, partout où il passe, le populisme se caractérise par sa totale incapacité à régler les problèmes qu’il dénonce.

Ma chronique pour L’Express 🔽

[LightOn @LightOnIO]

🇪🇺 LightOn rejoint le consortium AION porté par @Ardian, @Artefact_France, @BullAIDestiny, @Capgemini, @EDFofficiel, @GroupeIliad, @orange et @Scaleway pour contribuer à l’émergence d’une AI Gigafactory européenne en France.

Au sein du consortium, @LightOnIO apportera son expertise de l'IA en production : déployer, fiabiliser et passer à l'échelle des systèmes d'IA en environnement réel.

À l’ère des agents, l’IA devient une infrastructure stratégique.
Lire le communiqué 👇
https://t.co/KjaYdkwtwg

[LightOn @LightOnIO]

🎬 Les cas d'usage impossibles - Saison 1, épisode 02 : Juridique

Une crise d'approvisionnement secoue vos opérations.
📄 19 contrats fournisseurs à arbitrer. 🌍 5 langues. ⚖️ 3 systèmes juridiques.

Tous mentionnent "force majeure".
❓ Combien vous protègent vraiment ?

Ce cas d'usage ne teste pas la capacité d'une IA à retrouver des clauses de force majeure, il teste sa capacité à tenir face à l'ambiguïté juridique en entreprise.
Quand "trouver la clause" ne suffit pas, parce que l'annexe la contredit page 14.

📰 Lire l'analyse complète :
https://t.co/qHg93SFaIa

💻 Tester le scénario sur votre propre corpus documentaire : https://t.co/RNJQKEI7FA

[LightOn @LightOnIO]

🇱🇺 @LightOnIO at MeluXina alongside @luxprovide as part of the @DeployAIeu initiative.

@DeployAIeu is one of the first European projects aiming to build a commercial sovereign AI platform for end-to-end scalable solutions combining:

⚙ European compute infrastructure
🧠 Enterprise-grade generative AI
🔗 Interoperable deployment models

@LightOnIO and @luxprovide are already working on joint go-to-market foundations around sovereign AI infrastructure for European enterprises and institutions.

The stack is becoming real.

👉🏻 https://t.co/mHLXqka3sX

#DeployAI #SovereignAI #EuropeanTech #LightOn #LuxProvide #MeluXina

@AIonDemand

[IdePHICS lab @idephics]

Check out our lastest work on understanding deep learning! With @DandiYatin, @l_arnaboldi, Matteo Vilucchio and Hugo Tabanelli 👇👇👇

[Krzakala Florent @KrzakalaF]

What if a theory of deep learning could be built from iterated kernel spectral methods?

Feature learning, advantage of depth, emergence of concepts, convnets filters.... and a new backprop-free algorithm too! We have it all!

Introducing Neural LoFi 🧵
https://t.co/BvwyQhMLuR

[Krzakala Florent @KrzakalaF]

What if a theory of deep learning could be built from iterated kernel spectral methods?

Feature learning, advantage of depth, emergence of concepts, convnets filters.... and a new backprop-free algorithm too! We have it all!

Introducing Neural LoFi 🧵
https://t.co/BvwyQhMLuR

[Igor Carron @IgorCarron]

Introducing @LightOnIO 's Agent-ModernColbert: the more you use it, the more you save.

[LightOn @LightOnIO]

Reason-ModernColBERT topped BrowseComp-Plus with just 149M parameters.
Now, Agent-ModernColBERT adds ~10% on top.

Reaches GPT-5 + Qwen3-8B with GPT-OSS-120B.
Still 149M parameters.

Fully Open. Smaller. Cheaper.
Kudos to @antoine_chaffin for the work 👏🏻

Full benchmarks, methodology, model, data, and training code in the blog ↓
https://t.co/z9Ne0CfsL9

[bitwise @bitwise0X2A]

tested @LightOnIO's recent 149M denseOn inside kbolt (fast local retrieval engine) replacing the default 300M EmbeddingGemma on BEIR FIQA + SciFact.

- FiQA: nDCG@10 0.3695 → 0.4767, Recall@10 0.4218 → 0.5996.
- SciFact: essentially the same

latency wise denseOn was a bit slower possibly because Gemma's inference path is more mature and both have ~100M non-embed params?

[LightOn @LightOnIO]

@LightOnIO × @Dassault3DS

Some conversations signal where the market is heading.

On June 9, LightOn will be part of @outscale EXPERIENCES 2026 to discuss a shared vision of enterprise AI: sovereign, secure, and built for operational deployment at scale.

As AI moves into critical environments, infrastructure, governance, and execution can no longer be separated.
More soon.

📍 CNIT Forest, Paris La Défense

hashtag#SovereignAI hashtag#EnterpriseAI hashtag#OUTSCALEExperiences hashtag#LightOn hashtag#DassaultSystemes

[Antoine Chaffin @antoine_chaffin]

btw everything is open
The model, of course: https://t.co/IliqTUpUBA
But also all of the recipe, including data: https://t.co/oDuFuTaTMd

Beating these new benchmarks won't be easy, so come land an hand please

[LightOn @LightOnIO]

Retrieval is all about signal, architecture, not raw power.

🦾The late interaction LateOn 0.1B competing with the dense Gemini 2

[LightOn @LightOnIO]

From deploying agents → to owning your data, models, and infrastructure.

That shift is being explored today at @gosimfoundation Paris 2026, a key European event for open-source and sovereign AI.

🎤Today, @staghado took the stage for @LightOnIO :
“LightOnOCR: Pushing the Performance–Efficiency Pareto Frontier of Open OCR Models”

🔍 Parsing is the first step of any AI system
Before anything else, your data needs to be extracted, structured, and usable.

👉🏻 Discover LightOn orchestrated pipeline, from parsing to grounded answers: https://t.co/RNJQKEHzQ2

[Igor Carron @IgorCarron]

We solve search.

Be the first on the waiting list:
https://t.co/9GbukDqyYk