The AI community building the future.
External link for Hugging Face
We’re on a journey to solve and democratize artificial intelligence through natural language.
Paris, FR
Hugging Face reposted this
There are some wild apps at MCP 1st Birthday party!🤯 Interact with your agents using LIVE chats while the agents live in the virtual world interacting with environment! The Emergent Show - a 24x7 Live Show w/ Unreal Engine 5 + Model Context Protocol. The Emergent Show - You can even create your AI Agent's virtual avatar! Check out the app here: https://lnkd.in/gtjwb2EN 🔥 FINAL WEEKEND! Submissions close Nov 30, 11:59 PM UTC 🔥 https://lnkd.in/gYaMWfC9
Hugging Face reposted this
Benjamin Bossan and I set out to prepare a comprehensive presentation on a relatively untouched topic -- "Testing in ML Libraries". We poured our learnings from maintaining two widely used libraries (PEFT and Diffusers) while working on the presentation. Below is the overview: * Revisiting the existing topic of tests 🤷 * A bit about ML libraries and their kinds * Approaching tests for the OSS libraries at 🤗 * Practical concerns and how we address them (with concrete examples) Benjamin recently had a chance to present this in person at PyData Amsterdam. 📹 Find the full video: https://lnkd.in/gmkf9NSZ 🛝 Find the slides: https://lnkd.in/gyqfN5Yu
Hugging Face reposted this
Did we just copy the BEST feature of LinkedIn? Yes. Yes, we did.
Hugging Face reposted this
A community of developers just built a math-proof visualizer — and it only took them a weekend. Check out the Gradio app created as part of the MCP 1st Birthday Hackathon — it converts static math problems into interactive, step-by-step visual proofs. 👀 Want to join the fun? 🔥 FINAL WEEKEND! Submissions close Nov 30, 11:59 PM UTC 🔥 Win cash prizes + API credits 🔥 https://lnkd.in/gYaMWfC9 Join Huggingface Discord for live support: https://lnkd.in/g2JMSzg5
Hugging Face reposted this
Flux.2 is a BIG BOI! 🐘 The inference takes more than 80GB VRAM A small post on how one can run it on a NVIDIA L4 (22 GBs of VRAM)? 😉 The text encoder is Mistral AI's "Small-3.2-24B". It has 24B bf16 parameters which occupies 24*2=48 GBs VRAM. We can simply host it as an endpoint and delegate the encoding to our remote endpoint. This frees up 48 GBs worth of local VRAM One would quickly notice that text encoding here is a little nuanced, than input text, output hidden states. This calls for some custom logic and wrapping it as an API for people to use. I have built an end 2 end tutorial on just this. > custom text-encoding function > serve it with FastAPI > package it as a Docker container > deploy it on 🤗 Inference Endpoints Did you know, with the help of remote text encoding, cpu offloading, and quantization, you can now run Flux.2 in a colab notebook? Happy Flux Day to one and all! PS: Thanks to Sayak Paul and Erik Kaunismäki for all the guidance and help for this project.
Hugging Face reposted this
Introducing the HF upload speed Benchmark: We edit a single cell in a Parquet file. 🐢 git-lfs: 5.9 MB/s (re-uploads the whole file) ⚡ git-xet: 419 MB/s (smart deduplication) 📦 Now available on: macOS (brew install git-xet) Linux Windows (winget install git-xet) Stop letting network bandwidth slow you down. git-xet lets you push to Hugging Face hub at Nitro speeds. Kudos to XetHub team and Di Xiao!
Hugging Face reposted this
Finally, NanoChat has landed in transformers! 🚀 And we went wild on this deep dive blog post. 🔗 https://lnkd.in/e4w5gTPS We all went down the rabbit hole with Andrej Karpathy’s architecture, tinkering and learning from the ground up. But integrating it into the Hugging Face ecosystem opens up a new world of possibilities beyond just the model itself. In this deep dive, I explore the lineage of the architecture, the integration process, and the powerful tools you can now use with it. It includes: - detailed comparison of nanochat and canonical implementation. - explainer on how and why transformers user modularity. - deep dive examples on inference and training in torch, TRL, and vLLM. It was a lot of fun working on this, so I hope folk enjoy the read.
Hugging Face reposted this
You can now edit datasets directly on the Hugging Face Hub. The future of datasets for AI is here. 🔥 No more downloading a 500MB CSV just to fix 3 mislabeled rows. The workflow: 1. Spot an error in your dataset 2. Click edit in Data Studio 3. Fix the cells 4. Commit with a message 5. Repeat! Every change is versioned like code. Full traceability. The interesting part: collaborative curation. Your team can make commits to the same dataset, review each other's changes, and improve data quality together. I just published a blog post with a practical example: https://lnkd.in/d3JVvvh9 This will change how we maintain datasets for AI. Please leave a comment with your feedback and ideas to help us shape the future.
Hugging Face reposted this
🚀 We just shipped a big upgrade to our imitation-learning-in-simulation playground in Hugging Face LeRobot, built together with Lightwheel ! You can now teleoperate robots in sim (keyboard or real robot) and collect training demos instantly. This makes it possible to run real IL research on harder, more realistic manipulation tasks, even if you don’t have hardware. New tasks: 🟠 pick orange to the plate 🧺 fold cloth (yes!) 📦 pick 2 “e” toys to the box 🔴 lift red cube Through our partnership with Lightwheel, LeIsaac was integrated into EnvHub on day one, a best-practice integration that strengthens both ecosystems and pushes simulation-first robotics forward. 👉 Load a task via EnvHub, start teleop-ing, record your demo, upload the data on the hub and you’re ready to train or test. Get started here!: https://lnkd.in/dBpiUEKj
Hugging Face reposted this
Our new paper Economies of Open Intelligence is out and covered by Melissa Heikkilä at the Financial Times. This work gives the clearest picture yet of how global power is shifting inside the open AI ecosystem, and what that means for the open source community that has grown around it. Using 2.2B model downloads across 851k models, we trace how dominance in open AI has shifted dramatically over the last five years. US industry giants, once holding 40–60% of all downloads, have seen a steep decline. In 2025, for the first time, China’s model developers surpassed the U.S. in adoption, driven primarily by DeepSeek and Qwen. This shows the story of who dominates the global race for AI is not finished; it has just started! The global race is playing out not in closed labs, but in the open weight ecosystem where adoption and influence can be directly measured. At the same time, community developers and unaffiliated users have become the backbone of open AI. Industry once produced ~70% of downloaded models; today it’s closer to 37%. Community developers, online collectives, and individual contributors now shape more than half of the ecosystem. The emergence of a new layer of community intermediaries (quantizers, repackagers, adapter builders) shows how much innovation and practical usability now comes from outside large labs. This dual narrative is powerful: global AI power is shifting, and the open source community is increasingly the one steering the direction of the ecosystem. We’re releasing the paper, the full dataset, and a live monitoring dashboard because understanding these shifts shouldn’t be restricted to insiders; it should be a public resource for researchers, policymakers, and the community itself. Thanks for the incredible collaboration Shayne Longpre, Christopher Akiki, Campbell Lund, Atharva Kulkarni, Emily Chen ❤️ And as always, thanks to the amazing colleagues at Hugging Face who supported this work especially Lysandre Debut Avijit Ghosh, PhD Irene Solaiman Yacine Jernite