Julia-1: decision model that runs on almost anything (supersoniclabs.ia.br)

Supersonic Labs has unveiled Julia 1, a compact decision model that leverages the multilingual encoder mmBERT-small and is designed to operate on a range of devices, including CPUs. Julia 1, with 144.3 million parameters, evaluates questions against multiple answer options, successfully executing tasks such as classification, ranking levels, and yes-or-no decisions. In initial evaluations, the model demonstrated a correct decision rate of 73.15% over 2,000 typed decisions and significantly outperformed existing benchmarks in various tasks, although it struggled with long lists of similar categories, highlighting areas for improvement. The significance of Julia 1 lies in its ability to run efficiently on limited hardware while providing a versatile decision-making framework across 52 locales, including multiple languages. This innovation allows for further exploration in compact AI models, especially under constrained budget conditions—its development costs amounted to just over $100. The findings from Julia 1 will inform future iterations, such as Julia 2, which aims to develop a proprietary foundation architecture. Ultimately, Julia 1 offers a glimpse into the potential of budget-friendly, adaptable AI solutions, with plans to release access via an API soon.