The model that didn't exist, so you made it yourself
- Source
- huggingface.co
- Date
Context
A Hugging Face blog post from October 8, 2026 describes ML Intern, an in HuggingChat that plans a training job, asks for a budget before spending, runs a small test, then trains, evaluates and publishes a model on Hugging Face hardware. The author used it to build six models.
Reported results include a citrus-disease image model, a of Qwen3.5-2B on 3,017 annotated images, that rose from 14.9% to 52.8% accuracy on 335 test photos. A 260M-parameter image model to four steps scored 0.536 on GenEval, which tests whether generated images match their prompts, against 0.563 for its 50-step teacher. Total compute was about $103 in GPU and CPU job charges, excluding the author's time.
The post also reports limits: the citrus model still misjudged roughly half the test cases, one style model's look leaked into unrelated prompts from step 500, and some jobs failed on missing packages or wrong paths and were resubmitted. It is one author's showcase of six projects, not a controlled test of the agent.
- AI agent — An AI system that doesn't just answer once but works toward a goal in a loop — taking actions, reading the results, and deciding what to do next.
- fine-tuning — Taking a trained model and training it a bit more on your own examples so it gets better at one specific job.
- distillation — Training a small, cheap model to imitate a big one's outputs, keeping much of the capability at a fraction of the cost.
Checking sign-in…
Loading comments…