It cheapens the cost of fine tuning, it doesn’t make the model itself smaller at inference time.