Hugging Face’s summer was anything but quiet. Between June and August, the Hub jumped from 10,000 public model repositories to over 16,000 — a 60% increase in three months. That’s not just growth; it’s a signal that open-source ML is becoming the default way researchers and engineers share work.
The biggest launch is Spaces Beta, a free hosting service for machine learning demos built with Gradio or Streamlit. You can deploy a working app in minutes, set secrets, add custom requirements, or manage it straight from a GitHub repo. The community has already built over 100 Spaces — recipe generation with Chef Transformer, speech-to-text with HuBERT, video segmentation with DINO, and Paint Transformer for turning photos into paintings.
But the smaller features might matter just as much. TensorBoard integration gives every model repo an automatic, free TensorBoard instance if traces exist — public or private, any library. Model cards can now list evaluation metrics that auto-link to Papers With Code leaderboards, so sharing a model means instant side-by-side comparison with community results. And 18 browser widgets let users test models directly, including new ones for sentence similarity, audio classification, text-to-image, and object detection using Google’s ViT and Facebook AI’s DETR.
The community events show where the energy is. A JAX/Flax sprint in July drew nearly 800 participants who produced 170+ models, 22 datasets, and 38 Spaces demos with free TPUv3s from Google Cloud. One winning project, DALL·E mini, generates images from text prompts while being 27 times smaller than the original DALL·E. The first part of Hugging Face’s free online course also launched in June, covering Transformers, Tokenizers, Datasets, Accelerate, and the Hub. If Spaces catches on the way model hosting did, Hugging Face is positioning itself as more than a repository — it’s becoming the closest thing to an app store for open-source AI.
