From the source
A Hugging Face user describes building six custom models in a few days using ML Intern, an agent that plans, budgets, trains, evaluates, and publishes models on Hugging Face hardware.
Examples include a citrus disease VLM fine-tuned from Qwen3.5-2B that improved accuracy from 14.9% to 52.8% for about USD 1.90, a Huggy character LoRA for FLUX.2, and a camera-angle LoRA for Qwen-Image 2.1 costing about USD 16 in compute.
The post shares prompting techniques such as requesting a baseline before training and setting a strict spending limit.
