Adaptive, on-device playlists powered by a lightweight mini llm.

Feature description:

On-device AI-driven music assistant that creates personalized playlists based on your listening habits, without sending any data to the cloud. It learns directly from your music library and adapts to your preferences over time. Predicting similar songs based on recently played tracks. You can also type a short request like “Make me a happy playlist for summer” or “Chill beats for studying”.

The model needs to be initially trained for this purpose, but it could also continuously adapt based on user listening behavior.

On current smartphones, even 3billion models run pretty well locally. Much smaller models would probably be sufficient for this purpose.

Problem solved:

Currently, Playlists can only be created based on rules or randomness.

Brought benefits:

In my opinion, this would fill an missing feature gap.

Other application solutions:

 
Similar to FUTO Keyboard, which uses a 30-million-parameter mini Transformer model for text correction and prediction. Futo Keyboard LM Docs

A larger model would probably be necessary, but it would only run for a very short time. You could set how often it updates the playlist. On current smartphones, even 3b models run pretty well locally.

 

Additional description and context:

 

 

Screenshots / Mockup: