According to @emollick, a small local model produced transcription and full harmony analysis of a video of 4′33″, showcasing on-device music analysis.