This paper deals with direct data-driven design of model-reference controllers whose number of parameters is constrained. Input-output (I/O) sparse controllers are introduced and proposed as an alternative to low-order controller tuning. The optimal I/O sparse controller is shown to be never worse than the optimal low-order controller with the same number of parameters and a suited design procedure based on convex optimization is derived. The theoretical concepts are illustrated by means of a benchmark simulation example.