Building custom AI models for under $40 with autonomous agents

A developer recently demonstrated how to create specialized, small-footprint AI models using an autonomous agent called ML Intern. Over the course of a few days in October 2026, the engineer built six distinct models, ranging from agricultural disease detectors to image generation accelerators, with a total compute cost of approximately $103. The process relied on careful prompt engineering and strict budget controls within the Hugging Face ecosystem.

What happened

The project began with a specific need: a lightweight version of the prompt rewriter included in Qwen-Image 2.1. The official model is large, requiring about 20 GB of memory and significant processing time. Existing alternatives on the Hugging Face Hub were merely compressed versions of the same heavy model. To solve this, the developer tasked ML Intern, an agent available in HuggingChat, with creating a smaller, efficient alternative. The agent produced a 0.8 billion parameter model that runs on a CPU, uses only a quarter of the teacher model’s tokens, and achieves valid output 99.7% of the time. The entire process, including data labeling by the larger model, cost just $16.

Encouraged by this result, the developer created five more models using the same workflow. Each project started as a conversation in HuggingChat with ML Intern enabled. The agent handled planning, budgeting, testing, training, evaluation, and publishing. It operated on Hugging Face hardware, ensuring that every step was tracked and costs were managed. The resulting models were published publicly with detailed evaluation metrics in their model cards, demonstrating a repeatable pipeline for custom AI development.

How it works

The core of this workflow is structured prompting. The developer spent significant effort refining the initial instructions, which grew from 450 words to nearly 2,000 words over six projects. Each prompt begins with a clear statement of the goal and the rationale. It then specifies the dataset, base model, and training script. Crucially, the prompt includes a section labeled "Verified facts, do not re-derive," which lists known constraints and technical details. This prevents the agent from wasting resources rediscovering information the user already possesses, such as specific trainer capabilities or known issues with fallback tools.

Two specific instructions are vital for success. First, the prompt must request a baseline performance score before any training begins. This allows for a clear comparison to measure improvement. Second, it requires a smoke test—a small, low-cost trial run—to verify that the training process works before committing to the full job. For example, the agent might run 50 training steps and check if the model weights have changed. Finally, the prompt sets a hard spending limit, such as $12, and requires the agent to ask for permission before exceeding it. Since ML Intern starts with a zero-dollar budget, this control mechanism ensures costs remain predictable.

Key details

• The developer built six models, including a citrus disease detector, a character-specific LoRA, and a 4-step image generator.
• Total compute costs for all six projects amounted to approximately $103, with individual projects ranging from $1.90 to $37.
• ML Intern operates within HuggingChat and requires explicit permission to spend money, starting with a zero-dollar budget.
• The agent automates dataset creation, training, evaluation, and publishing to the Hugging Face Hub.
• Prompts include a "Verified facts" section to prevent redundant research and enforce technical constraints.
• Baseline metrics and smoke tests are mandatory components of the prompt to ensure quality and cost efficiency.

Why it matters

For software engineers and ML practitioners, this approach lowers the barrier to creating custom models. Traditionally, fine-tuning requires deep expertise in infrastructure management, hyperparameter tuning, and data pipeline construction. By delegating these tasks to an agent, developers can focus on defining the problem and curating the data. The ability to produce specialized models for under $40 makes experimentation accessible to individuals and small teams who cannot afford large-scale GPU clusters.

Moreover, the emphasis on small, efficient models addresses a growing need for edge deployment and cost-effective inference. The 0.8 billion parameter rewriter, for instance, runs on a CPU, making it suitable for applications where latency and hardware constraints are critical. This shift towards distillation and specialized LoRAs allows developers to tailor AI capabilities to specific domains, such as agriculture or brand-specific image generation, without relying on generic, resource-heavy foundation models.

What you can do

• Access ML Intern via HuggingChat and enable the agent mode to start experimenting with automated model building.
• Review the example prompts provided by the developer on GitHub at yvrjsharma/ml-intern-prompts to understand effective structuring.
• Start with a small budget cap, such as $10, to test the agent’s capabilities on a simple task before scaling up.
• Include a request for baseline metrics in your prompts to ensure you can quantify the improvement of your fine-tuned model.
• Define a smoke test procedure, such as a short training run, to verify the pipeline before committing to full-scale training.
• Share your resulting models on social media and tag relevant communities to contribute to the open-source ecosystem.