---
title: "The model that didn't exist, so you made it yourself"
description: "The author wanted a small version of the prompt rewriter that ships with Qwen-Image 2.1, but the official 9B model requires about 20 GB of memory. Using ML Intern in HuggingChat, they created a 0.8B version that runs on a CPU, returns valid output 99.7% of the time, and uses about a quarter of the teacher’s tokens. The compute for the entire project, including having the 9B model label 8,797 example requests, came to USD 16. Over a few days, the author made five more models using the same approach, each starting as a message in HuggingChat with ML Intern switched on. The author’s prompts include verified facts, a baseline request, a smoke test, and a spending cap. The six models built include a citrus disease vision model, a Huggy character LoRA, a camera-angle LoRA, a doodle-in LoRA, a pocket rewriter, and a distilled text-to-image model."
---

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

[Hugging Face](https://yomu.fyi/company/hugging-face) · yuvraj sharma, Abubakar Abid · Oct 8, 2026

**Type:** Tutorial

## Summary

The author wanted a small version of the prompt rewriter that ships with Qwen-Image 2.1, but the official 9B model requires about 20 GB of memory. Using ML Intern in HuggingChat, they created a 0.8B version that runs on a CPU, returns valid output 99.7% of the time, and uses about a quarter of the teacher’s tokens. The compute for the entire project, including having the 9B model label 8,797 example requests, came to USD 16. Over a few days, the author made five more models using the same approach, each starting as a message in HuggingChat with ML Intern switched on. The author’s prompts include verified facts, a baseline request, a smoke test, and a spending cap. The six models built include a citrus disease vision model, a Huggy character LoRA, a camera-angle LoRA, a doodle-in LoRA, a pocket rewriter, and a distilled text-to-image model.

## Context

The author wanted a small version of the prompt rewriter that ships with Qwen-Image 2.1, but the official model is a 9B model requiring about 20 GB of memory. They used ML Intern to create a 0.8B version that runs on a CPU, returning valid output 99.7% of the time and using about a quarter of the teacher’s tokens, for a total compute cost of USD 16.

## Approach / What changed

The author describes a workflow where they describe a desired model in a prompt to ML Intern, an agent in HuggingChat. ML Intern plans the work, requests a budget, runs a small test, then trains, evaluates, and publishes the model on Hugging Face hardware. The author includes verified facts, a baseline request, a smoke test, and a spending cap in each prompt.

## Takeaways

- ML Intern can fine-tune a 2B vision model on 3,017 annotated citrus disease images for about USD 1.90, improving accuracy from 14.9% to 52.8% on a test set of 335 photos.
- A camera-angle LoRA for Qwen-Image 2.1 was built by rendering 24,722 transparent images of household objects at 24 angles each, training for 2,000 steps on one A100 for about USD 3.75, with the whole project costing about USD 16.
- The pocket rewriter project distilled a 9B teacher model into a 0.8B student that runs on a CPU, using 8,797 generated image requests and costing USD 16 in total compute, with the student returning valid output 99.7% of the time.

## FAQ

### How do I ensure the model is actually improving during training?

The author includes a line asking for a baseline score before training, such as the base model's zero-shot score on the same metric. They also include a smoke test, like running 50 training steps and checking that the saved weights changed, before paying for the full run.

### How do I control the compute cost when using ML Intern?

The author caps total spend in the prompt, e.g., 'Cap total spend at USD 12 and ask me before exceeding it.' ML Intern begins every task with a zero-dollar budget and needs permission before executing paid jobs, so the spending limit is strictly enforced.

### How can I prevent ML Intern from wasting budget on re-deriving known information?

The author includes a section titled 'Verified facts, do not re-derive' in the prompt. This section lists information the author has already checked, such as which trainer supports transparent images or which GitHub issues make a fallback trainer risky, so the agent spends its budget on the work instead of rediscovering known facts.

**Tags:** [AI Agents](https://yomu.fyi/topic/ai-agents), [Fine-Tuning](https://yomu.fyi/topic/fine-tuning), [Hugging Face](https://yomu.fyi/topic/hugging-face), [LLMs](https://yomu.fyi/topic/llm), [LoRA](https://yomu.fyi/topic/lora)

- Source: [Hugging Face](https://huggingface.co/blog/building-with-ml-intern)
- Source URL: https://huggingface.co/blog/building-with-ml-intern
- Ingested by Yomu: 2026-10-09T00:00:18.139Z

[Read original post](https://huggingface.co/blog/building-with-ml-intern)