Frequency content is an important characteristic of earthquake shakings. The mean period (Tm) of a ground motion has been regarded as the preferred frequency content parameter. In this paper, a new predictive model for Tm is developed based on the horizontal components of ground motions selected from the expanded NGA-West2 database. The new model includes a piece-wise linear moment magnitude (Mw) term with a breaking Mw value as 5 and a cutoff value as 7.3, which is adopted based on the distribution of empirical data as well as physical considerations. A trilinear rupture distance term with breaking points at 100 km and 200 km is also adopted in the model. The average shear wave velocity at top 30 meters (Vs30) and a basin depth parameter (Z1) are used to represent the effect of local site conditions on Tm. A forward directivity term that takes effect at rupture distances within 40 km is included in the functional form. Besides, we derive a magnitude-dependent aleatory variability model to capture the varying within-event standard deviations versus Mw. The proposed model can be used to estimate Tm for earthquake scenarios with magnitudes in the range of 3–7.9 and rupture distance up to 300 km.