Recent advancements in artificial intelligence have predominantly focused on scaling large language models (LLMs), with major corporations like OpenAI, Google, and Meta developing models comprising hundreds of billions of parameters.This emphasis has fostered a well-organized research community dedicated to optimizing these expansive “black-box” models, extensively discussing both their successes and limitations. Conversely, the systematic exploration of how to utilize these black-box models—through approaches such as prompt engineering and ensemble techniques, which leverage LLMs without retraining—has received comparatively less attention, and the attention it has received is often poorly organized.In this position-paper, we aim to demonstrate why the field of small language models deserves significantly more focus. We outline the unique challenges facing research in this area, pose key questions related to the development and application of small LLMs, and highlight gaps that persist in the current literature.