Rice leaf diseases cause yield losses of up to 30% in smallholder farms, yet most detection models are too heavy for the low-cost devices farmers can afford. We propose a distilled vision transformer with 3.1 million parameters that runs at 24 frames per second on a single-board computer. Trained on 18,400 field images from four provinces, the model reaches 94.6% accuracy across six disease classes, within 1.2 points of a full-size baseline. Field trials with 42 farmers show that diagnosis time fell from days to under a minute.