Julia 1: A 144M-Parameter Decision Model Trained for $104

Brazil's Supersonic Labs has released a decision model named Julia 1, with an impressive claim. It's a 144.3-million-parameter model that operates on a CPU, a tablet, or a 12-year-old laptop-class chip, and the entire training cost is a mere US$104.08. Developed by the small AI lab, Julia 1 is designed to pick the correct answer from a list of options, making it suitable for tasks such as classifying support tickets or determining the urgency of an issue.

Unlike traditional AI pipelines that generate text, Julia 1 focuses on providing labels for specific questions. It can handle three types of questions: choosing one option from a closed list, placing an input on an ordered scale, or returning the probability that a statement is true. This approach eliminates the need for parsing answers, reducing the likelihood of producing malformed responses.

The model was trained using mmBERT-small, a multilingual encoder from Johns Hopkins CLSP, and then adapted to a decision-making task. The training cost was a relatively low R$540, or US$104.08, thanks to utilizing existing infrastructure and avoiding expensive GPU resources.

Julia 1 has shown promising results in various benchmarks, including typed decisions, AG News, and emotion classification. However, it struggles with longer lists of similar categories and lacks external knowledge or multi-step reasoning capabilities. The model's impressive performance comes at a low cost, making it an attractive option for businesses looking to implement narrow AI solutions on a budget.