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

A Hugging Face blog post published last week describes how its author used ML-intern, an agent available in HuggingChat, to create six models over a couple of days, each starting as a chat message and ending as a published model on the Hub with its evaluation in the model card. ML-intern plans the work, requests a budget before spending, runs a small test before a full job, then trains, evaluates and publishes on Hugging Face hardware.

The first project, Pocket Rewriter, was a small replacement for the prompt rewriter shipped with Qwen-Image 2.1. The official rewriter is a 9B model needing roughly 20 GB of memory, so the author generated 8,797 example image requests, had the 9B teacher rewrite them on one A100 in 2 hours 37 minutes (about USD 6.50), filtered to 1,840 training examples, and trained 0.8B and 2B students in 12 and 18 minutes on an A10G (USD 0.75 for both). The 0.8B student returns valid output 99.7% of the time, uses about a quarter of the teacher's tokens, runs on a CPU, and ships as an 812 MB GGUF file.

Other builds include a citrus-disease vision model fine-tuned from Qwen3.5-2B on 3,017 annotated images across 21 classes; on 335 test photos it named the right problem 52.8% of the time after two epochs on one A10G, against 14.9% for the base model. A character LoRA for Huggy was trained on 84 captioned drawings on FLUX.2 klein base 4B, with checkpoints saved every 100 steps. A camera-angle LoRA for Qwen-Image 2.1 used 1,030 scanned objects rendered at 24 angles each; training ran 2,000 steps in about 90 minutes on one A100. A doodle-in LoRA built 6,042 training pairs and a 160-pair test set; at 6 steps with the Viggle turbo LoRA, 67.5% of objects were detected where drawn, at 4.7 seconds per edit. Agate-Preview-002-4step distilled a 260M-parameter text-to-image model from 50 steps down to 4, exporting it to ONNX for browsers; GenEval improved from 0.509 to 0.536 after a second run, against the teacher's 0.563 at 50 steps.

Compute costs ranged from about USD 1.90 for the citrus model to about USD 37 across both Agate runs, with the camera-angle LoRA, doodle-in LoRA and Pocket Rewriter each around USD 16 or USD 24. The author advises prompting with verified facts, asking for a baseline score before training and a smoke test with a check, and capping spend, since ML-intern starts with zero budget and needs permission before paid jobs. The seven prompts used are on GitHub.

Why it matters: developers can produce task-specific models from a chat prompt with small compute budgets, and the resulting students run on CPUs or in browsers rather than requiring the larger teacher models.