Every single robot coming out of Hugging Face is aesthetically astounding in a different way. This is peak performance. This is the future of open-source

At a time when people are understandably focused on the daily chaos in Washington, these articles describe the rapidly accelerating impact that AI is going to have on jobs, the economy, and how we live. axios.com/2025/05/28/ai-…

Today, we are introducing SmolVLA: a 450M open-source vision-language action model. Best-in-class performance and inference speed! And the best part? We trained it using all the open-source LeRobot datasets in the Hugging Face hub! But how? 🫳🏀

I have bittersweet news to share. Yesterday we merged a PR deprecating TensorFlow and Flax support in transformers. Going forward, we're focusing all our efforts on PyTorch to remove a lot of the bloating in the transformers library. Expect a simpler toolkit, across the board.

KerasHub is a collection of over 70 popular pretrained model architectures -- LLMs, VLMs, image generation models, etc -- that work with JAX, TF, PyTorch. They all support HuggingFace checkpoints -- you can load any HF model with them for the corresponding architecture.

If any Irish speakers are interested in helping annotate some of the BLEnD examples to Irish, please let me know (can be as little as a few examples). We aim to have Irish included in the next release. huggingface.co/datasets/nayeo…

"The great unbloating" of transformers continues. Over the past few weeks, 10+ PRs were merged, aiming to simplify code across the library. This brought in refactors for Attention, the Cache, a new linter. We're improving type hints everywhere, and are checking type checkers.

We have finally released the 📝paper for 🥂FineWeb2, our large multilingual pre-training dataset. Along with general (and exhaustive) multilingual work, we introduce a concept that can also improve English performance: deduplication-based upsampling, which we call rehydration.

We are so excited to announce a new open-source challenge in collaboration with Proxima Fusion : unlocking fusion with AI If you haven't followed, fusion is how the sun make energy and is –in the long term– our best bet on a clean, safe, and virtually limitless energy In the

I hear wildly conflicting reports about the impact of the number of CCDs on Epyc CPU bandwidth. Does anyone have a range of CPUs available to measure? If so, how many CCDs do you actually need to get the theoretical ~500GB/s socket bandwidth? Do access patterns matter?

New thrill sport: Mushroom foraging where you don't have an identification key, you just show photos of them to a language model. Disagreeing with the LM's assessment is STRICTLY against the spirit of the sport

My latest post: The American DeepSeek Project Build fully open models in the US in the next two years to enable a flourishing, global scientific AI ecosystem to balance China's surge in open-source and an alternative to building products ontop of leading closed models.

I asked Deepseek about RoPE because I realized I never understood why you split the vectors into pairs of values before rotating, and it told me to go try it with triples instead and to have fun with the gimbal lock hell of non-commutative 3D rotations I was about to experience