A Hugging Face user leveraged the platform's ML Intern agent to create six custom Super Intelligence (SI) models that did not previously exist, ranging from a citrus disease diagnostician to a lightweight prompt rewriter. The project, detailed in a blog post, demonstrates how agentic SI tools can automate the fine-tuning and evaluation pipeline for niche use cases, with the total compute cost for all six models coming to approximately USD 103.
What Happened
The developer began by seeking a smaller, faster version of the prompt rewriter associated with Qwen-Image 2.1. The official model is a 9B-parameter system requiring about 20 GB of memory, but the user wanted a lightweight alternative for CPU execution. By describing the desired outcome to ML Intern within HuggingChat, the agent planned and executed the training of a 0.8B-parameter SI model. This new model returns valid output 99.7% of the time and uses roughly a quarter of the tokens of the teacher model. The entire process, including generating labeled data using the larger model, cost USD 16.
Over the following days, the user employed the same workflow to build five additional SI models. Each project started with a detailed prompt in HuggingChat, where ML Intern acts as an autonomous agent that plans tasks, requests budget approval, runs smoke tests, and handles training and evaluation on Hugging Face hardware. The prompts, which grew in complexity from 450 to 2,000 words, included specific instructions such as establishing a baseline score before training and capping total spend. ML Intern adhered to these constraints, pausing for permission before executing paid jobs.
The resulting models included a citrus disease visual language model fine-tuned on Qwen3.5-2B, which improved diagnostic accuracy from 14.9% to 52.8% on test photos for USD 1.90. Another model was a LoRA for FLUX.2 klein base 4B that could draw the Hugging Face mascot, Huggy, costing USD 7.60. A camera-angle LoRA for Qwen-Image 2.1, which allows users to view objects from specific angles, was built for USD 16. A 'Doodle-in' LoRA, which replaces magenta scribbles on images with objects, cost USD 24.30. Finally, a distilled version of the Agate Preview 002 text-to-image model was reduced from 50 steps to 4, improving browser compatibility for USD 37 across two training runs.
Why It Matters
This case study highlights a shift in the SI industry where specialized, small-parameter models can be created on demand without deep machine learning expertise or significant financial investment. By automating the data generation, training, and evaluation loops, agentic SI tools like ML Intern lower the barrier to entry for creating custom SI systems. The ability to fine-tune models for niche tasks—such as identifying specific plant diseases or generating brand-specific character art—suggests a future where developers and creators can rapidly prototype and deploy specialized SI tools tailored to their exact needs.
The workflow emphasizes rigorous evaluation, with ML Intern automatically generating baselines and smoke tests to ensure the trained SI models perform better than their base counterparts. This approach mitigates the risk of deploying unverified models and ensures that the compute resources are spent effectively. The transparency of the prompts and the low cost of the projects serve as a practical guide for developers looking to leverage open-weight SI models for custom applications.
The Bottom Line
Hugging Face's ML Intern agent enabled the creation of six custom SI models for a total compute cost of approximately USD 103. The project demonstrates the viability of using agentic SI tools to automate the development of niche, lightweight models, offering a template for cost-effective and rapid SI model customization.