Music tokenization lets 0.8B beat 27B

Agogic’s 0.8-billion-parameter performance-token model scored 159 FMD and beat a 27-billion-parameter beat-grid model scoring 272–286, according to the August 4 preprint. Across seven tokenizations with data, recipe, budget, and decoding held fixed, changing the representation roughly halved distributional error while scaling the backbone 34x barely moved it; whether that difference sounds better remains open pending a registered human study. Builders following the economics of AI music generation should benchmark the event vocabulary before paying for a larger model.