Welcome fastai to the Hugging Face Hub

Hugging Face and fast.ai have quietly closed a gap that's existed for years: there was no easy way to publish a trained fastai Learner somewhere other people could actually grab and use. Now there is. With the huggingface_hub library installed (pip install huggingface_hub["fastai"], plus fastai 2.4 or newer), a single call to push_to_hub_fastai uploads a full Learner — model, dataloaders, loss function, the works — to a Hub repo, complete with an auto-generated model card and full git-based version control.

The demo they walk through is the classic cats-vs-dogs classifier from the first chapter of the fastbook, trained in six lines and pushed to a repo called espejelomar/identify-my-cat. Pulling it back down is just as trivial: from_pretrained_fastai loads the repo id and you're predicting again, no retraining required. In their test it nailed a cat photo with 100.00% confidence, which is either a great model or a very obvious cat.

The more interesting example, though, is the Blurr integration. Blurr lets fastai users tap into Hugging Face Transformers directly, so the post trains a DistilBERT-based sentiment classifier on the IMDB sample dataset using Blurr's high-level API, then pushes that to fastai/blurr_IMDB_distilbert_classification on the Hub. That's a small but telling detail: it shows fastai's ecosystem and Hugging Face's transformer models aren't separate worlds anymore, they can share a single pipeline and a single hosting layer.

None of this is flashy — there's no new architecture, no benchmark chart. It's plumbing. But plumbing is what turns a library full of enthusiastic hobbyists and course-takers into a place where trained models actually circulate instead of dying on someone's laptop. Given how many people started their ML careers through Jeremy Howard's courses, as one of the Hugging Face writers here did, lowering the friction to share what they build seems like a genuinely useful, if unglamorous, move.