Hugging Face's ML Intern builds six small models in days on a tiny budget
In a Hugging Face blog post, author yvrjsharma describes using the ML Intern agent inside HuggingChat to build six models in a couple of days, including distilling Qwen-Image 2.1's 9B prompt rewriter into a 0.8B 'Pocket' version that runs on CPU (99.7% valid output, about a quarter of the teacher's tokens) and fine-tuning Qwen3.5-2B for citrus-disease diagnosis, which rose from 14.9% to 52.8% accuracy on 335 test images. The agent plans the work, requests a budget before spending, runs a smoke test, then trains, evaluates and publishes to the Hub; the author puts the first project's total compute at about USD 16. The post also shares prompting advice, such as asking for a pre-training baseline and capping total spend.
It shows how agentic tooling lets an individual cheaply distill large models into small, device-friendly ones.