This post is from the WWDC26 Audio Q&A.
Is it appropriate to use the Foundation Models framework on-device to power a natural language sound recommendation feature (e.g. user types 'I can't focus' → model returns a frequency/playlist recommendation from a predefined set)? Or is this use case better served by a custom Core ML model? What's the recommended on-device inference approach in iOS 27 for this type of structured recommendation with a small output space?