
AI @LightonIO Language Models, Information Retrieval, Knowledge Distillation PhD
Paris, France
Joined May 2020
- Tweets1.4K
- Following946
- Followers1.2K
- Likes5.7K
Pinned Tweet
Releasing ColGREP and LateOn-Code models 🚀
ColGREP is a multi-vector search tool built in Rust made for coding agents. It's an hybrid grep which supports both grep features and semantic retrieval. Run 100% locally.
You get two SOTA code retrieval model within ColGREP
Raphaël Sourty retweeted
Replying to @paulomouraj @LightOnIO
The model performs incredibly well!
It scores lower than mDenseOn on my internal French benchmark where I'm certain there are no leaks. But in the Sparse category, it outperforms the Sparseup approach by more than 4 pts 🤯
I'm going to read your blogpost very very carefully 👀
Super valuable-information-dense blogpost on a super promising method. Congratulations!
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/l…
Blog: huggingface.co/blog/lightona…
SOTA models and SOTA blogpost 🔥
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/l…
Blog: huggingface.co/blog/lightona…
Raphaël Sourty retweeted
I mean, when you know the authors you already know it's going to be insane, but honestly the blog post was so interesting
Stellar work, and very timely: the future is sparse, and getting strong sparse models easily with training SAE on top of dense/LI models is the way
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/l…
Blog: huggingface.co/blog/lightona…
Raphaël Sourty retweeted
Just sparsify everything! Insane release by Paulo and the blog post is a delightful read, give it a go!
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/l…
Blog: huggingface.co/blog/lightona…
We release SlideOn and mSlideOn at @LightOnIO today, two sota sparse retrievers, models designed by @paulomouraj and I.
Those models are really strong for first stage retrieval, you can add then a multi-vector model as a reranker and then a decision model 👇
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/l…
Blog: huggingface.co/blog/lightona…
We build those models from sota embeddings models we released earlier this year. We give a lot of details to the recipe in the HF blogpost available in @paulomouraj thread
Those results are obtained thanks to an heavy work on the data as well which we released earlier this year
Raphaël Sourty retweeted
« Ce n’est que du matériel, ils ont juste à faire jouer les assurances, pas de quoi en faire un drame. »
Il faut n’avoir absolument aucune conscience de ce qu’est être commerçant ou entrepreneur pour dire un truc pareil.
Voir son outil de travail partir en fumée, c’est traumatisant. Voir ses économies investies dans un commerce qui met parfois des années à rapporter un maigre salaire, c’est violent. Beaucoup ont en plus été frappés, blessés, humiliés.
Les assurances ? Si les petits commerces arrivent à se faire rembourser une partie des pertes, ce sera déjà une victoire. Elles ne couvriront pas tout, si elles couvrent quelque chose, et ce sera un enfer administratif et judiciaire pour beaucoup de victimes, ajoutant au traumatisme.
Ce n’est pas « que du matériel ». Ce sont des vies brisées, des violences physiques et psychologiques, qui auront des conséquences.
Rien ne justifie cela. Aucun motif politique. Il n’y en a pas, ce sont juste des pillages, tout le monde le sait. Ceux qui essayent de minorer ou de justifier cette violence (qui touche d’ailleurs des gens souvent modestes, dont la vie va être encore plus difficile désormais) s’en rendent complices.
Raphaël Sourty retweeted
🚨 Tested @LightOnIO's ColGrep-Agent (2B local code-search model) on our own repo across 3 coding agents.
Same 22 questions, same model, OFF vs ON, 3 runs each:
• pi: 43% fewer tokens
• OpenCode: 37%
• Claude Code: 21%
ACCURACY UNCHANGED.
The model needs ~3 GB of VRAM. 🧵
Raphaël Sourty retweeted
Our next generation of open weights multimodal decision models comes in two sizes, pplx-decider-v1.2 and pplx-decider-flash-v1.2.
Text and vision scores improve. Inference is up to 5x faster: all questions in a request run in one pass.
Flash is the top 4B on the Decision Index.
Raphaël Sourty retweeted
Today, we're releasing the next iteration of our decider models.
On the novelty plate:
- Two sizes, 27B and 4B, depending on your needs
- A new attention schema to enable better batching
- Better results (please read more later!)
Raphaël Sourty retweeted
Just received the award for the #SISAP2026 Indexing Challenge here in Brno! 🏆🇨🇿
Very proud of the tusKANNy team (@fmnardini, @cosimorulli1, @rventurini_). We took home both 1st and 2nd place in Task 3 (Learned Sparse Retrieval)! 🧵👇
#VectorSearch #InformationRetrieval
Raphaël Sourty retweeted
Today we are releasing Scout 1.0, giving frontier performance at a fraction of the cost! Have a look at the thread of Ameet for more details 🧵
Raphaël Sourty retweeted
For those of you who need a quick summary of all the sudden findings across so many disciplines, here is a summary of all the "And they fall..." posts. Enjoy!
Raphaël Sourty retweeted
Huge thanks to my co-authors @Blyzi1, @froger_romain, and @zehavoc for doing such an amazing job, especially @Blyzi1 and @froger_romain for leading this work!!
📄 Paper: arxiv.org/abs/2610.06064
💻 GitHub: github.com/Blyzi/trustmi
🤗 Hugging Face: huggingface.co/TrustMI/
Raphaël Sourty retweeted
This release is super exciting, so I had Opus-5.5 make me a small demo of @LightOnIO's ColGrep-Agent-2B + @pleiasfr's Pleias-RAG-1B for coding queries related to the stanford-futuredata/ColBERT codebase. I made a 17-minute video walking through my 5 test queries (for each one, I also had Opus 5.5 generate a response with code snippets as a comparison).
My demo's repo: github.com/vishalbakshi/colg…
🤗 Model: huggingface.co/lightonai/col…
Congrats and big thanks to the LightOn team for blessing us with lightweight, open, intelligent, code localization! This is absolutely essential as we head towards a (hopefully) local-powered open future.
---
Apologies for the low-quality captions in my video ("LightOn" --> "Lydon" lol). It's done using a vibe-coded voice dictation app using whisper.cpp's small-en model that I have not yet fine-tuned because I am still collecting training data.
Raphaël Sourty retweeted
Wow... 😲
A new 2B local model claims it can cut your Claude usage by 34%!
ColGrep-Agent-2B does the code search, then hands Claude only the files and lines it needs.
That means fewer search turns and less usage. They say accuracy stays about the same, but this needs to be verified imo.
It's just 2.5 GB, and your code never leaves your machine for the search. It works with Codex and OpenCode too.
Looks interesting enough for me to try it out!
Raphaël Sourty retweeted
The Vector Search landscape is full of complex methods. At tusKANNy, we like to simplify things through abstractions.
Meet RerankIndex: a general framework for two-stage search, now available in kANNolo 0.12 (Python & Rust) 🧵👇
#VectorSearch #InformationRetrieval
Raphaël Sourty retweeted
New open retrieval model: UIBK-Embedding-v1 👀
179M parameters. A dense retriever and a ColBERT retriever in one model, one index, one search.
🥇 Best model under 1B parameters on BEIR
🥇 #1 on NanoBEIR among models up to 6B
Blog: huggingface.co/blog/abdoelsa…


























