Hugging Face's ML Intern trains a real model from a chat prompt
Now live in HuggingChat: it picks a model, builds a dataset, trains on a budget you set, ships weights, data and a demo to the Hub. One run cost under $0.50.
Copy markdown
Chat in, trained model out
Built on Hugging Face's open ml-intern agent, ML Intern now runs inside HuggingChat at huggingface.co/chat. Describe an idea in plain English and it picks a base model, builds a dataset, launches the training job, tracks it on a live dashboard, then pushes the trained weights, dataset, a report, and a working demo Space to the Hub.
It quotes the bill before it spends it
The agent estimates compute cost up front and holds to a budget cap you set, so an experiment can't silently drain your account. Hugging Face's demo run trained for about six hours and finished for under $0.50.
Build this weekend: a custom model, no training code
Point it at a narrow task — sort support tickets by urgency, flag defects in product photos, tag your own text — and you get back a fine-tuned model plus a shareable demo, with no PyTorch, CUDA, or GPU setup. The bar drops from 'ML engineer' to 'can describe the problem.'
Elsewhere: the same agent is open source
Prefer to self-host? huggingface/ml-intern is on GitHub — an open agent that reads papers, trains models, and ships them — so you can run the full loop on your own compute and keys instead of the hosted chat.