Deep learning models have achieved the state of the art in blood glucose (BG) prediction. However, most existing models can only provide single-horizon prediction and face a variety of real-world challenges, such as lacking hardware implementation. In this work, we introduce a new deep learning framework, the edge-based temporal fusion transformer, for multi-horizon BG prediction, and implement the trained model on a customized wristband with a system on a chip (Nordic nRF52832) for edge computing. On a clinical dataset with 12 people with diabetes, it achieved the smallest mean root mean square error for 30 and 60-minute prediction horizons.