A model's complexity is its capacity to fit varied shapes in the data. It is tuned through hyperparameters: depth of a tree, k in k-NN, degree of a polynomial, regularization. Too low, the model underfits; too high, it overfits. You look for the level that generalizes best.