The aim of this letter is the construction of a new model order reduction algorithm generalized to the multi-input/multi-output systems model order reduction. It is essentially based on the complete order model dominant modes retention. This involves the conversion of the overall significant information contained in the original complete order system into the reduced order approximant, permitting the construction of our approximant denominator. The approximant numerator is computed by means of genetic algorithm (GA) tools and a square error criterion. As a result, an optimal approximant of lower order is derived. To show its performances, to highlight some important of its characteristics, and to conclude to its efficiency, a comparative study is carried out. A numerical example is given, where our approximant model is compared to reduced order models computed from two recent and important techniques based on GA tools, namely, the stability equation and the modified pole clustering techniques.