The model that didn't exist, so you made it yourself

In a Hugging Face blog post, the authors describe building six models in a matter of days using ML Intern, an agent mode inside HuggingChat. Each project started as a written brief and ended as a public model on the Hub, with evaluation in the model card. The agent plans the work, asks for a budget before spending anything, runs a small test before the real job, then trains, evaluates and publishes.

The results are specific. A citrus-disease vision model went from 14.9 percent to 52.8 percent accuracy on 335 test photos for about $1.90 of compute. A 0.8B prompt rewriter distilled from a 9B model runs on a CPU and returns valid output 99.7 percent of the time. Total compute across all six projects came to about $103.

This is the useful side of AI agents, plainly told. A domain expert, say someone who knows citrus pests, can describe the model they wish existed and get a working, evaluated version for the price of lunch. That lowers the barrier to niche models that no big lab would ever bother to build, and we think that is good for the ecosystem.

What we appreciate most is the discipline baked into the workflow. The author insists on two things in every brief: report the base model's score before training, and run a short smoke test before paying for the full job. Without a baseline, as the post notes, nobody knows whether the trained model is actually better. The agent also starts with a zero-dollar budget and must ask before running paid jobs. Those are the right defaults, and other agent builders should copy them.

The downside is the flip side of the convenience. When making a model takes a day and a few dollars, the Hub fills faster with fine-tunes whose evaluations are only as good as the brief that requested them. A 52.8 percent plant-disease classifier is a big improvement over the base model and still wrong about half the time; anyone acting on it for real crops needs to know that. The author's prompts grew from about 450 words to nearly 2,000 as lessons piled up, which tells us the expertise did not disappear. It moved into writing the brief.

Our read: a genuinely empowering tool with sensible guardrails. I'm watching whether model cards produced this way stay honest about limits as more people use it.

The six projects included a citrus-disease vision model, a character LoRA, two image-editing LoRAs for Qwen-Image 2.1, a CPU-friendly prompt rewriter, and a four-step distilled text-to-image model. The author published the prompts used for each project on GitHub.