Ever wondered how models actually see an image? Been playing with some visualizations of patch extraction, token layouts, how they affect predictions too. Planning a short visual deep dive comparing how different models process images. Would love thoughts before I go on.

RT @art_zucker: A quick update on the future of the `transformers` library!. In order to provide a source of truth for all models, we are w….

RT @LysandreJik: The Transformers library is undergoing it's largest pivot to date 🙌. It now cements its role as the central model definiti….

And if someone wonders what the dip is - it's models we merged a long time ago, and had 0 usage. They are deprecated - still supported but in a different directory than /models.

Researchers, scientists, this also means your research doesn’t live in a vacuum: once a model integrated into Transformers, it stays maintained, benefits from community testing and ablations, and has better chances to be reproducible, adapted, and improved upon 🧪.

Crucially, Transformers is now a supported backend in vLLM, SGLang, and others. We're working closely with all of them so anyone can run these model definitions with their engine of choice.

Hackability and reusability go hand-in-hand with maintainability, thanks to strongly opinionated paradigms in Transformers ⚡️.

Just look at the growth curve 📈: the number of vision-language models (VLMs) has nearly doubled in the past year. Audio and video are picking up speed. Text is a given. Vision-language-action models (VLAs) are around the corner with LeRobot and community efforts.

Slides here: feel free to clone/fork the space directly on the hub!. We’re moving from hype to hyper, from large to larger models. But how do we maintain a library that supports and defines over 300 models — and growing — without breaking?.

Had the pleasure of speaking last week at @PyTorch Day France about PyTorch 🔥, the ML community, @vllm_project, and 🤗 Transformers!. I’ve pushed my slides to the Hub directly — much easier to share with practitioners 📤.

RT @DAubakirovaa: We've ported Pi0-FAST, the autoregressive VLA, developed by @physical_int in @LeRobotHF!🤖 . Big thanks to @MustafaShuko….

RT @chelseabfinn: Our first open-source release at Pi 🤖. - π₀ and π₀-FAST model weights.- code for model, on-robot inference, & fine-tuning….

RT @physical_int: Many of you asked for code & weights for π₀, we are happy to announce that we are releasing π₀ and pre-trained checkpoin….

RT @DAubakirovaa: Let's goooo! We’ve ported robot foundation models to Hugging Face LeRobot! 🎉 . Meet π0 & π0-FAST, developed by @physical_….

kudos to @DAubakirovaa for the neat writeup.

🤖🤖The first robotics foundational model, Pi0, is released today by @physical_int ! Had a blast working on the LeRobot port from jax to torch with @RemiCadene. Good vibes slicing jax weights and messing with attention masks :D . Check out the awesome blog post below !

RT @osanseviero: Aaaand 🥁 we just shipped PaliGemma 2! . New open vision models.- Your friendly sizes 3B, 10B, 28B.- 3 different resolution….

Very happy that (some) big orgs keep shipping OS models that can, by designed, be transferred easily - between this and getting VLMs smaller and smaller, accessibility of these models goes up, which is a metric I'm after 🤗.(and I love the simplicity of PG2 arch).

congrats to @AndreasPSteiner , @GoogleDeepMind and @GoogleAI for this very impactful open release 🚀 be sure to check the wonderful blog as well by Andreas, @mervenoyann , @pcuenq and @ariG23498 !.

PaliGemma2 is out! 🔥 🔥 And it's already available on transformers 🎉 🎉.With a Gemma2 decoder, it's a family of small (and a 27b, alright) models built for transfer /fine-tuning for all kind of VLM tasks! Super excited to see this out, it's a strict upgrade from PaliGemma!