Hugging Face

We're partnering with PyTorch, Hugging Face on the OpenEnv Challenge! Use Unsloth AI for RL & OpenEnv to win $10K in HF credits! As part of the University of California, Berkeley AgentBeats Competition, there is a special track just for reinforcement learning! If you can: 1. Create an RL environment, and publish to the HF Hub 2. Publish training notebooks with Unsloth, HF 3. Write a blog on HuggingFace Then submit an entry! You also get to publish a PyTorch blog! The AgentBeats competition details are at https://lnkd.in/gNDkHMnb For the OpenEnv track details, see https://lnkd.in/gZ93sqdd

2026 will be the year for AI for Science (and we're hiring for that!) But 2025 was where we laid the foundations: datasets, benchmarks, tooling, partnerships, and (most importantly) community. A few things I’m proud of from the past year: 🚀 We shipped and supported 20+ collaborations and releases, including: - OpenFold3 (from the OpenFold Consortium) - AQCat25 (a major catalyst dataset with SandboxAQ) - the Antibody Developability Challenge (with $60,000 in prizes from Ginkgo Datapoints) - and the ConStellaration dataset + challenge with Proxima Fusion + a steady stream of releases with teams at NVIDIA, IBM, and AI at Meta 📣 On the communications side, I spoke at World AI Summit (a panel + two fireside chats with the incredible Ben Kompa and Max Welling). We published blogs and resources for building across fusion, pharma data, antibody modeling, and more. There were papers like LeMat-GenBench (with a Hugging Face Space) and LeMat-Synth at NeurIPS (with the extremely good team at Entalpic). 🤝 There was “AI for Scientific Discovery is a Social Problem” (with Avijit Ghosh, PhD), which sparked a lot of scientists to reach out with real needs, ideas, and collaboration offers. 🛠️ Tooling-wise, with Quentin Lhoest and great open-source contributors, we’ve been pushing dataset support for scientific formats on the Hub (think FASTA, CIF, and other common bio/materials file types). I’ve also been supervising work to enable streaming and slicing large HDF5/HDF datasets, which matters a lot for physics-scale data (led by Mike Smith). 👯♂️ Community has been a big theme. After a viral post kicked things off, we launched Hugging Science: the Hub org is now ~800 members, the Discord is ~7,000 members (huge thanks to Adam Molnar), and there are multiple ongoing research projects there spanning topics like benchmarking PDEs, AI for allergies, and atomic simulation. 🔬 We also ran and supported a set of hackathons across microscopy, chemistry/materials, women’s health & longevity, cancer research, and protein folding. 🎉 One of my favorite moments: we hosted a Hugging Science event at NeurIPS— ~600 attendees, ~2,000-person waitlist, and a line around the block. Weeks later, someone recognized the limited merch in a totally different city and came over to talk science. That’s the kind of energy that this ecosystem is forming! 🔮 Looking to 2026: Hugging Science will be getting its own site (with dedicated resources for scientists) and we’re growing the team (EMEA and US roles), and we'll do stuff we can't even imagine yet! 🚀 If you want to be part of it: reach out, DM, or check out our careers page! If you know a friend, who would be a great fit, also let me know 🙏

Qwen3-VL has all you need for e2e multimodal RAG 🔥 I have put together a notebook for this, you can run on a free Colab (T4)! the models are 2B, we remove them once we're done embedding and reranking, saving memory 🙌🏻 if you have long documents though, make sure to use FA2 and quantize using bitsandbytes try it here https://lnkd.in/d8SiCByh

👾 `hf-mem` is all you need to estimate the required VRAM for inference of any model on Hugging Face based on Safetensors metadata. - Written in Python - Lightweight, only depends on httpx - Runs w/ Astral `uvx` as `uvx hf-mem --model-id ...` - Works with any Safetensors repository - Output inspired by U.S. Graphics Company TR-100 Machine Report

AI is accelerating AI research. On every Hugging Face paper page, we've deployed a chatbot (powered by open models such as GLM-4.6) to quickly understand papers by answering questions, summarizing key ideas, and providing context as you browse the latest research Next step, thanks to MCP HuggingChat goes full "AI Scientist mode" 🔥 Hat tip Mishig Davaadorj Niels Rogge Ahsen Khaliq and team

Hugging Face est heureux d’être sponsor du hackathon multi-écoles Hack'N'Roll organisé par Thales, un événement dédié à des challenges concrets de cybersécurité opérationnelle, au cœur des enjeux industriels et souverains. Soutenu par l’État dans le cadre de l’AMI « Train Cyber Expert » du Programme France 2030, ce hackathon vise à former les futurs experts cyber en les confrontant à des problématiques réelles, complexes et critiques.   Grâce au Hugging Face Hub, les participants auront la possibilité de : 🔓 s’appuyer sur des technologies ouvertes et auditables, adaptées aux environnements sensibles 🧠 accéder à des modèles IA pour l’analyse, la détection et l’automatisation 🤝 collaborer facilement entre étudiants, chercheurs et industriels 📚 favoriser la reproductibilité, la transparence et le partage des connaissances   À travers ce soutien, Hugging Face s’inscrit dans une dynamique collective visant à mettre l’IA open source au service de la sécurité numérique, de la formation et de l’écosystème académique. Ces initiatives illustrent le rôle que peut jouer l’Open Source pour accompagner l’innovation tout en répondant à des contraintes opérationnelles fortes. cc Penelope GITTOS

Hugging Face and Clem Delangue 🤗, what's the plan for cat-proofing Reachy? See my concept art below for my guess as to what would happen if I were to procure one. I have five cats, two big dogs, two children, and cattle if you would like some durability testing done.