Hugging Face Blog Details Six AI Models Built by ML Intern Agent for About USD 103

Hugging Face published a blog post detailing how its ML Intern agent executed the training and publishing of six specialized AI models in a few days, with a total compute cost of approximately USD 103. The agent plans tasks, requests budget approval, runs smoke tests, and publishes models with evaluations directly to the Hub based on user prompts.

Key Models Built

• Citrus Disease VLM: Fine-tuned Qwen3.5-2B on 3,017 annotated images. The model improved accuracy from 14.9% (base) to 52.8% on a 335-photo test set. Compute cost: USD 1.90.
• Huggy LoRA: A LoRA for FLUX.2 klein base 4B trained on 84 drawings to generate the Hugging Face mascot. Compute cost: USD 7.60.
• Viewpoint Orbit LoRA: A camera-angle LoRA for Qwen-Image 2.1 trained on 24,722 rendered images of household objects. Compute cost: USD 16.
• Doodle-in LoRA: A LoRA for Qwen-Image 2.1 that replaces magenta scribbles with objects, trained on 6,042 synthetic pairs. It achieved 67.5% object detection accuracy at 4.7 seconds per edit. Compute cost: USD 24.30.
• Pocket Rewriter: Distilled the 9B Qwen-Image-2.1-PE-T2I prompt rewriter into 0.8B and 2B student models. The 0.8B version runs on CPU and returns valid output 99.7% of the time. Compute cost: USD 16.05.
• Agate 4-step: Distilled the 260M-parameter Agate Preview 002 text-to-image model from 50 steps to 4 steps. It achieved a GenEval score of 0.536 (vs. teacher's 0.563 at 50 steps) and was exported to ONNX for browser use. Compute cost: USD 37.

Workflow and Cost

The author emphasizes that effective prompting involves defining verified facts, requesting baselines before training, and setting strict budget caps. The ML Intern agent enforces these limits, asking for permission before exceeding them. The total cost for all six projects, including data generation and training, was about USD 103.