Since the early 1990s, mismatch negativity (MMN) has been widely used to investigate sensory memory traces and predictive processing of musically relevant acoustic features, including timbre, relational pitch, and rhythm. Evidence consistently demonstrates that the human brain possesses innate musical abilities, which can be further refined through training. The interplay between innate predispositions and experience is supported by robust associations between MMN parameters and behavioral listening skills in both musically trained and untrained children and adults. These findings align with the interpretation of MMN as a prediction-error signal reflecting the precision of internal predictive models: MMN amplitudes decrease with increasing contextual complexity and show more frontal scalp distributions in musicians. As stimulation paradigms have become increasingly ecologically valid, distinguishing low-level sensory prediction errors indexed by MMN from higher-level frontal prediction errors has grown more challenging, supporting an integrative perspective in which generative models dynamically merge bottom-up sensory input with top-down priors across hierarchical levels of the central nervous system. In this review, we summarize evidence for a “musical intelligence” of the automatic MMN mechanism, highlighting its role within widespread cortical predictive coding processes.