Raphaël Sourty

I'm thrilled to announce the release of FastPlaid ! 🚀🚀 FastPlaid is a high-performance engine for multi-vector search, built from the ground up in Rust (with the help of Torch C++)⚡️ You can view FastPlaid as the counterpart of Faiss for multi vectors.

🚀 We’re excited to introduce Qwen3-235B-A22B-Thinking-2507 — our most advanced reasoning model yet! Over the past 3 months, we’ve significantly scaled and enhanced the thinking capability of Qwen3, achieving: ✅ Improved performance in logical reasoning, math, science & coding…

Great excuse to share something I really love: 1-Lipschitz nets. They give clean theory, certs for robustness, the right loss for W-GANs, even nicer grads for explainability!! Yet are still niche. Here’s a speed-run through some of my favorite papers on the field. 🧵👇

🚀 Introducing Qwen3-MT – our most powerful translation model yet! Trained on trillions of multilingual tokens, it supports 92+ languages—covering 95%+ of the world’s population. 🌍✨ 🔑 Why Qwen3-MT? ✅ Top-tier translation quality ✅ Customizable: terminology control, domain…

🚀 Introducing GLiClass‑V3 – a leap forward in zero-shot classification! Matches or beats cross-encoder accuracy, while being up to 50× faster. Real-time inference is now possible on edge hardware. huggingface.co/collections/kn… #TextClassification #NLP #ZeroShot #GLiClass

Ok the solution might be way easier than expected ... it's been a long time since we did not release a SOTA model, isn't it? 😇

Bye Qwen3-235B-A22B, hello Qwen3-235B-A22B-2507! After talking with the community and thinking it through, we decided to stop using hybrid thinking mode. Instead, we’ll train Instruct and Thinking models separately so we can get the best quality possible. Today, we’re releasing…

Some of the ModernBERT team is back with new encoder models: Ettin, ranging from tiny to small: 17M, 32M, 68M, 150M, 400M & 1B parameters. They also trained decoder models & checked if decoders could classify & if encoders could generate. Details in 🧵:

Introducing ColQwen-Omni, a 3B omnimodal retriever that extends the ColPali concept of multimodal retrieval with late interaction to audio chunks and short videos, with no performance degradation on visual document retrieval wrt our best models! (1/N)

TorchDR 0.3 is here with some major improvements, taking the library to the next level! TorchDR leverages vectorized implementation on GPU for super fast dimensionality reduction. Thanks to all the contributors!! Description below🧵

✊ Transformers... Assemble! Introducing ♊Ettin Suite, a SoTA open recipe to outperform existing Generative & Retrieval Models. Developed by @JohnsHopkins in collaboration with @LightOnIO, Ettin is the first-ever SoTA suite of paired encoder & decoder models. The revolution…

To anyone wondering what's the difference between encoders and decoders on downstream tasks when both models are trained the same way, this blog post is made for you. Very interesting resource and new models available, impressive work 🙌

🤔 Have you ever wondered how good ModernBERT is compared to decoders like Llama? We made an open-data version of ModernBERT and used the same recipe for encoders and decoders. Turns out, our encoder model beat ModernBERT and our decoder model beats Llama 3.2 / SmolLM2 🤯 🧵

The #SIGIR2025 Best Paper just awarded to the WARP engine for fast late interaction! Congrats to Luca Scheerer🎉 WARP was his @ETH_en MS thesis, completed while visiting us at @StanfordNLP. Incidentally, it's the fifth Paper Award for a ColBERT paper since 2020!* Luca did an…

Alright that’s it WARP coming to PyLate soon™️

This was a really enjoyable and approachable blog post. I think this is my favorite explanation of MaxSim and I'm going to use it moving forward. But this snippet doesn't do the whole post justice---read the whole thing!

Awesome new lib from @bclavie to make MaxSim computation quicker on CPU ! In the same vibe, also need to underline pylate-rs by @raphaelsrty ! Both libs go a long way into making Late Interaction (ColPali, ColBert) ever more accessible !