A novel control-relevant model reduction technique for nonlinear systems is proposed utilizing the idea of the balanced truncation. Unlike the widely-accepted Karhunen-Loeve method where the state basis for the reduced system is found from the state snapshots, the proposed technique takes into account the input, state, and output information together and provides a near-balanced reduced-order model that approximates the system map instead of the state snapshots. Performance of the technique is demonstrated for a linear system and a non-adiabatic fixed-bed reactor model.