During the initial training of a motor sequence, performance becomes progressively faster but also increasingly reproducible and consistent. However, performance temporarily becomes more variable at mid-training, reflecting a change in the motor representation and the eventual selection of the optimal performance mode (Adi-Japha et al., 2008). At the cerebral level, whereas performance speed is known to be related to the activity in cerebello-cortical and striato-cortical networks, the neural correlates of performance variability remain unknown. We characterized the latter using functional magnetic resonance imaging (fMRI) during the initial training to the Finger Tapping Task (FTT), during which participants produced a 5-element finger sequence on a keyboard with their left non-dominant hand. Our results show that responses in the precuneus decrease whereas responses in the caudate nucleus increase as performance becomes more consistent. In addition, a variable performance is associated with enhanced interaction between the hippocampus and fronto-parietal areas and between the striatum and frontal areas. Our results suggest that these dynamic large-scale interactions represent a cornerstone in the implementation of consistent motor behavior in humans.