MAGIC is a training-free, plug-and-play post-merging calibration framework that improves merged models by calibrating layer-wise magnitudes (in feature space, weight space, or both).