I trained a local on-device (3B) medical note model and benchmarked it vs frontier models (results + repo)