The primary motor cortex (M1) is the most prominent motor-output area of the cerebral cortex. In M1, layers 2/3 (L2/3) and 5a (L5a) constitute intermediate layers upstream of layer 5b (L5b), the major motor-output layer. During motor learning, the microcircuits of M1 are thought to self-organize to integrate various types of signals. However, it remains unknown how the neuronal activities of L2/3 and L5a of M1 are reorganized during learning of a motor task. We conducted two-photon calcium imaging in the mouse M1 during two-week training sessions of a self-initiated lever-pull task (requiring forelimb use). We quantified the movement information carried by the neurons as the mutual information between predicted and recorded lever trajectories. In L2/3, the accuracy of neuronal ensemble prediction of lever trajectory remained unchanged globally throughout the training period. However, in L5a, the ensemble prediction accuracy steadily improved and one-third of neurons evolved to contribute substantially to ensemble prediction in the late stage of learning. The L2/3 network may represent coordination of signals from other areas throughout learning, whereas L5a may participate in the evolving network representing well-learned movements. The distinct dynamic networks in L2/3 and L5a would be core elements that drive the L5b motor output for well-learned movements.