We are releasing 📄 FinePDFs: the largest PDF dataset spanning over half a billion documents! - Long context: Documents are 2x longer than web text - 3T tokens from high-demand domains like legal and science. - Heavily improves over SoTA when mixed with FW-EDU&DCLM web copora.

This kind of evals are very interesting, and I wish there were a lot more I'm not in the team that thinks a high score is actually good (because it goes against instruction following), but it's great for monitoring and understand the different flavor of post-trained models. It's

This reply from Karen is spot on. AI discourse has effectively devolved into mindless chatter because of the anchoring to a shared mythology - everyone (boosters, doomers, even some critics) is endlessly debating & dissecting a version of the technology that doesn't even exist.

I’m thrilled that our INTIMA benchmark, developed to study how AI models handle companionship-like interactions, was featured in Forbes last week. Nothing is quite as amazing as seeing your work not only used but also picked up by journalists to reach a wider audience. The

🚀 Life update: I’ve joined 🤗Hugging Face as AI Scientist & Educator, starting a new track on **Mechanistic Interpretability of LLMs** 🧠🤖 Over the past 7 years at Ubisoft 🎮, I explored how AI, science & gameplay intersect. I worked on cutting-edge LLM-powered NPCs,

🤖 As AI-generated content is shared in movies/TV/across the web, there's one simple low-hanging fruit 🍇 to help know what's real: Visible watermarks. With others Hugging Face , I've made sure it's trivially easy to add this disclosure to images, video, chatbot text. See how:

SmolDocling just got a HUGE improvement, meet GraniteDocling!🚀 Improved performance in all the ways that matter: multilingual, more reliable, but still tiny at 258M params!🤏 It's lightning fast, process a page in 0.35 sec on a consumer GPU using < 500MB VRAM⚡

Updated the evaluation guidebook with a new deep dive! 2025 panorama of all the important and next level evaluations that you need to know to build *actually impactful and useful* models! (Assistant tasks, games, forecasting, and more) Tell me wyt! :) github.com/huggingface/ev…

By far the most concise and informative guide on post-training evals I've seen in a long time - highly recommended reading!

Clémentine is continuing to do some of the most important work on evals in the world <3

when you give up on this nebulous idea and illusion of prestige, you will finally find peace and freedom. submit to TMLR and JMLR.

Check out The Washington Post's awesome audit of Sora by Nitasha Tiku! washingtonpost.com/technology/int… They cite the Data Provenance Initiative and quote our Joanna!

Reuters just reported that Meta will soon use generative AI interactions to target ads across Facebook and Instagram. That’s exactly the kind of shift we explore in our blogpost: 👉 Advertisement, Privacy, and Intimacy: Lessons from Social Media for Conversational AI with

New paper 📜: Tiny Recursion Model (TRM) is a recursive reasoning approach with a tiny 7M parameters neural network that obtains 45% on ARC-AGI-1 and 8% on ARC-AGI-2, beating most LLMs. Blog: alexiajm.github.io/2025/09/29/tin… Code: github.com/SamsungSAILMon… Paper: arxiv.org/abs/2510.04871

Come say hi and get some copies of the SmolLM3 Blueprint! 🤗

Very cool paper! You can discuss with the author here: huggingface.co/papers/2510.04…

The Hugging Face Science team is in Montreal for COLM 2025 🍁! Rumour has it, limited-edition tees are up for grabs for top Hub contributors, ranging from smolLM to XL. 😉

So proud to see Reachy Mini named one of the Best Inventions of 2025 by TIME! Huge credit to the Pollen Robotics and Hugging Face teams, turning a concept into thousands of units sold and shipped in under 6 months. We might not be as slick as some other robotics companies (we

I spoke with MIT Technology Review about one of the hardest design questions in conversational AI: should an AI ever be allowed to hang up on a human? In the piece by James O'Donnell, we discussed how cutting users off can be harmful when strong emotional bonds or dependencies have formed.

We just released the beta version of the open-source software for Reachy Mini! It means that anyone, thanks to the amazing Google DeepMind mujoco simulation platform, can start building Hugging Face spaces, datasets and models, even if you haven't received your robot yet.