Human activity recognition (HAR) is essential for applications such as healthcare monitoring, fitness tracking, and smart environments, yet deploying accurate and interpretable models on resource-constrained devices remains challenging. In this paper, we propose XTinyHAR, a lightweight, transform…