Hugging Face

We've just surpassed 3 million models on the Hub 🚀 the community is accelerating towards an open, distributed future where AI is open everywhere, for everyone

With the recent model releases, we have been seeing that model checkpoints alone aren't enough. Model publishers are also shipping kernels to ensure these models run reasonably. A good example of this is the recent LTX-2.5 model from Lightricks. They used neighborhood attention kernels (a.k.a NATTEN) in the Video VAE to efficiently decode the video frames. We packaged the NATTEN kernels and published them on the Hub, making them very easy to use WITHOUT regressions. I think this pattern will become increasingly common in future model releases as well.

Are current open OCR models good enough to unlock historical knowledge? Daniel van Strien and I have been working on this question for the last few months — and today we are proud to introduce FineBooks, a new collaboration between Hugging Face and EleutherAI. Our first release is the BHL OCR Leaderboard: 14 open OCR models scored against 2,165 historical book pages transcribed by expert researchers from six volumes of the Biodiversity Heritage Library — in English, French, German and Latin. https://lnkd.in/eguACYSR The ground truth behind the leaderboard was produced in 2011–12 by the IMPACT project and BHL-Europe, released under CC-BY, and preserved by IMPACT's technical manager Clemens Neudecker on Github after both projects ended. This is what open science is all about: releasing your data openly so that down the line someone can re-use it for purposes you never planned for. Exactly the same open science and open data philosophy is at the core of FineBooks. The next step: we'll re-run OCR across the roughly 200,000 public domain volumes from the BHL and release the improved text openly as the first FineBooks dataset.

Meta released Muse Glimmer 30B: multimodal model for your Claw/Pi setups 🔥 we tested and fine-tuned the model for you, and shipped day-0 support in transformers and llama.cpp, including DFlash for 2-4x speed-ups 🥵 read our blog https://lnkd.in/dCzPcyhz

Training Agents with RL takes datasets, environments, training frameworks, and models. All of those are built and maintained (best) by the open source community. We held our first RL meetup at the @reflection office in SF this week alongside builders from Databricks, Hugging Face, Reflection, NVIDIA, Meta, Prime Intellect, Fleet, Modal, Mercor, and more. We focused on how we build and integrate high quality Rl Environments, with two topics: - Zach Wentz (Reflection) auto-validating agentic environment with scalability and security - Adithya S Kolavi (Hugging Face) on using OpenEnv as a runtime on top of Harbor in training, including multi-harness setups (a la Kimi K3) Beyond the talks, RL practitioners shared what they're seeing across the ecosystem, and several new RFCs came directly out of the room. That's what happens when you get this much diversity of thinking in one place. For context: We started OpenEnv back at Meta in partnership with Hugging Face and Unsloth and launched it at the PyTorch Conference last October. The vision was simple: build a lingua franca for RL environments, and the whole ecosystem builds to one interface so models generalize faster. Monday was the first time I saw that vision sitting in a single room and I’m excited for many more.. Here is the OpenEnv Repo if you’d like to contribute 👉 https://lnkd.in/g8_HK3H6 Cheers! P.S.  A special shout out to Zach Wentz and Adithya S Kolavi for giving amazing talks, Ben Burtenshaw for co-organizing and Becca Schwabe, Angie Ruiz-Mitchell and Helen T. for all of your support to make this happen. Cheers!