Small Language Models in Healthcare: Capabilities, Challenges, and Future Directions

Large Language Models (LLMs) exhibit significant proficiency in various natural language processing tasks. Nonetheless, their direct implementation in healthcare is limited by significant computational expenses, concerns regarding data privacy, and restricted interpretability. In contrast, Small Language Models (SLMs), generally comprising fewer than 8 billion parameters and often less than 3 billion in clinical applications, present a more pragmatic and privacy-aware option. This paper offers an overview of SLMs in healthcare, focusing on architectural innovations, training strategies, and domain-specific adaptations that facilitate their implementation in clinical environments. This work examines healthcare applications in decision support, documentation, and patient interaction, emphasizing technical adaptations including multimodal integration and knowledge distillation. Besides, the paper summarizes risk–benefit considerations to facilitate responsible adoption and proposes an updated taxonomy that classifies models based on deployment readiness and modality support. Finally, the paper concludes by identifying significant challenges and proposing future research directions for the reliable and scalable integration of SLMs in practical healthcare settings.