Hugging Face ML Intern: How One Developer Built Six Custom Models for About $103 (and the Prompt Playbook Behind It)

Hugging Face has published a hands-on account of ML Intern, an agent inside HuggingChat that takes a plain-language request and carries it all the way to a published model on the Hub. In the October 8, 2026 post "The model that didn't exist, so you made it yourself", authors Yuvraj Sharma and Abubakar Abid describe building six models in a few days for about USD 103 of compute in total, and shares the prompt structure that kept each run on track.

The interesting part is not the agent hype. It is the specific workflow: a budget the agent cannot exceed, a baseline before training, a cheap smoke test before the expensive run, and a model card with the evaluation attached. This post walks through what was built, what each project cost, and which of those habits you can reuse with any coding or training agent.

TL;DR: what Hugging Face reported

What is ML Intern, in plain terms?

Think of it as a junior ML engineer you brief in chat. According to the post, ML Intern plans the work, asks for a budget before spending anything, runs a small test before the real job, then trains, evaluates and publishes on Hugging Face hardware. Each of the six projects started as one message in HuggingChat with ML-intern switched on and ended as a public model with its evaluation in the model card.

That matters because the hard part of a custom model is rarely the training loop. It is assembling a dataset, choosing a base model, picking a trainer that supports your data type, catching silent failures, and knowing whether the result beat what you started with. The post shows an agent doing that glue work, including resubmitting jobs that failed on missing packages or wrong paths.

If you are new to the underlying ideas, our guides on what fine-tuning is and when to use it and what AI distillation is cover the vocabulary used below.

The six projects and what each cost

The author groups the work into four ideas: a model that knows your field, a model that draws your character, a model that does a new trick, and a model that fits your device.

1. Citrus Doctor: a model that knows your field (about USD 1.90)

A general vision model can describe a yellowing citrus leaf, but telling a mite problem from a magnesium deficiency, plus remedies, is much harder. The author assembled a training set merged from three sources hosted by the Project-AgML organization on the Hub, producing citrus-disease-vlm-instruct: 3,017 annotated images across 21 pests, illnesses, nutritional gaps and treatments.

ML Intern fine-tuned Qwen3.5-2B on it, after benchmarking the base model first. On 335 test photos the base model named the right problem 14.9 percent of the time. After two epochs on one A10G, the fine-tuned model reached 52.8 percent. Compute cost was about USD 1.90. The post notes the first prompt, about 450 words, did not even include the verified-facts section described below, and the model still more than tripled the base accuracy.

2. Huggy LoRA: a model that draws your character (about USD 7.60)

Image models know many characters, but not Huggy in the flat style of Hugging Face brand assets. The author asked for a LoRA on FLUX.2 klein base 4B, trained on 84 captioned drawings. The agent saved a checkpoint every 100 steps and rendered the same prompts with each, which made picking one easy. Step 200 was the first where Huggy was fully on-model; from step 500 the style bled into prompts that had nothing to do with Huggy. The LoRA also works on the distilled klein model at 4 steps.

3. Viewpoint Orbit and Doodle-in: models that do a new trick (about USD 16 and USD 24.30)

For the camera-angle LoRA, the author noticed that nobody had built one for Qwen-Image 2.1 a few days after its release. ML Intern rendered 1,030 scanned household objects from Google Scanned Objects at 24 angles each (24,722 transparent images) on a CPU job costing a few cents, finalized 461 objects for training and 40 for testing, and trained over 23 camera instructions. Training ran 2,000 steps in about 90 minutes on one A100 (around USD 3.75). The whole project took roughly half a day and 48 jobs, about USD 16.

Doodle-in is a LoRA where you draw a magenta scribble on a photo, name an object, and the model swaps the scribble for that object while preserving lighting and composition. No dataset existed, so the prompt described how to build one: take an Open Images photo, remove one object with the LaMa inpainting model, draw a scribble where it was, and use the untouched photo as the target. The agent built 6,042 training pairs and a 160-pair test set, 40 of which came from 23 object classes held out of training, and recorded the author and license of every source photo. After comparing checkpoints on 48 test pairs it picked step 500. In the post's evaluation, 67.5 percent of objects were detected where they were drawn, at 4.7 seconds per edit; unseen classes landed about as reliably as seen ones (65.0 versus 64.2 percent). Cost: about USD 24.

4. Pocket Rewriter and Agate: models that fit your device (about USD 16 and USD 37)

The motivating example: the official prompt rewriter that ships with Qwen-Image 2.1 is a 9B model needing about 20 GB of memory that thinks for thousands of tokens before writing a paragraph. The author wanted a small version. ML Intern generated 8,797 short image requests, had the 9B teacher rewrite them on one A100 (2 hours 37 minutes, roughly USD 6.50), filtered down to 1,840 examples, and trained 0.8B and 2B students in 12 and 18 minutes on an A10G. The 0.8B student returns valid output 99.7 percent of the time with about a quarter of the teacher tokens and ships as an 812 MB GGUF for CPUs.

Agate Preview 002 from Logolabs is a 260M-parameter text-to-image model small enough for a browser, but it needs 50 steps with guidance, about 100 network passes per image. ML Intern distilled it to 4 passes in two runs. The first cached 155,000 training images as latents, baked guidance into the model, and cut steps in stages from 16 to 8 to 4. The 4-step student beat the teacher run at the same 4 steps on GenEval and FID, then was exported to ONNX for the browser. A second run on 24,000 more teacher pairs lifted GenEval from 0.509 to 0.536, against the teacher's 0.563 at 50 steps. Total across both runs: about USD 37.

The cost table

Per the post, these are GPU and CPU job charges reported for each session. They do not include the author's time or the cost of the chat model doing the planning, so treat them as the compute floor rather than an all-in price.

The prompt playbook: what the author says to include

The post is candid that the first message is where the effort goes. The first prompt was about 450 words; by the sixth project it was closer to 2,000, because each project taught something worth carrying forward. The structure, as described:

1. The idea in one line, and why you want it.
2. The exact pieces: the dataset, the base model, the training script.
3. A "Verified facts, do not re-derive" section. Anything already checked goes here so the agent spends budget on work, not rediscovery. For the camera-angle LoRA it listed which trainer had just added transparent-image support and which open GitHub issues made the fallback trainer risky.
4. A baseline request. The citrus prompt asks for the base model's zero-shot score on the same metric before training, so the gain is visible. Without it you get a model and no idea whether it is better.
5. A smoke test with a check attached. For the image LoRAs: 50 training steps, then verify the saved weights actually changed, before paying for the full run.
6. Deliverables and a spend cap. Define what belongs in the model card, and state something like "Cap total spend at USD 12 and ask me before exceeding it."

The author's point about the budget is that ML Intern starts at zero and needs permission before paid jobs, so the cap is enforced rather than aspirational. If you leave it out, the agent proposes a couple of paths by project size and asks which you prefer.

These rules generalize beyond Hugging Face. Any agent that can launch paid jobs benefits from a baseline, a cheap canary run, and a hard cap. We make a related argument in our agent loop design guide on self-correction and memory.

Why this matters for builders

Small, specialized models are getting cheap to make. Single-digit dollar runs that triple accuracy on a narrow task (Citrus Doctor) change the build-versus-prompt decision. Our fine-tuning versus RAG decision guide is a good companion for deciding when that is worth it.

Distillation is the repeatable pattern. Two of the six projects (Pocket Rewriter and Agate) are teacher-student compressions where the student wins on the metric that matters for deployment: latency, memory, or step count. That echoes the small-model trend in VibeThinker-3B and the tiny image models like Moebius 0.2B for inpainting, and the edge focus of Liquid AI's open d1 decision models.

Dataset building is where agents help most. The Viewpoint Orbit and Doodle-in datasets were synthesized from existing assets (scanned objects, inpainting) with provenance recorded. Data is also what Hugging Face invests in directly, as in The Stack v3 with 5 trillion tokens.

Caveats worth keeping in mind

• These are the author's reported numbers. The post is written by a Hugging Face team member about a Hugging Face product, with results on small, author-chosen test sets (335 photos, 48 pairs, 160 pairs). Treat them as demos, not benchmarks.
• Narrow wins. A 0.8B rewriter that mimics a 9B teacher on one prompt distribution is not a general replacement for the teacher.
• Failures are part of the cost. The Viewpoint Orbit run needed 48 jobs and Doodle-in 59, including resubmissions after missing packages and wrong paths. Compute charges include such reruns, but your debugging attention does not.
• Licensing still applies. The Doodle-in pipeline logged the license of every source photo, which is the habit to copy when you build datasets from public images.
• Check what you publish. ML Intern pushes public models to the Hub. Review the model card and dataset before sharing anything built on private or restricted data.

How to try it

1. Open HuggingChat and switch on ML-intern mode.
2. Pick a model you wish existed and a dataset you have, or one you can describe in a paragraph.
3. Start from the author's free example prompts on GitHub and adapt the structure above.
4. Set a small cap (the post's runs ranged from about USD 2 to USD 37), require a baseline and a smoke test, and read the results before raising the cap.
5. Compare against a plain prompted baseline of a bigger model before you decide the fine-tune was worth it.

For more agent-driven workflows and small-model coverage, browse the explainx.ai blog.

Related reading

• What is fine-tuning an LLM: complete guide
• What is AI distillation
• Fine-tuning vs RAG: grounding decision guide
• Designing Claude agent loops: self-correction and memory
• VibeThinker-3B: small model, frontier reasoning
• Moebius 0.2B image inpainting

Details reflect Hugging Face's October 8, 2026 post and may change as ML Intern evolves; check the original for current pricing and features.