We investigated the source scaling of earthquakes (Mw 4.6-8.9), mostly from the Taiwan orogenic belt, and made a global compilation of source parameters to examine the scaling self-similarity. Finite-fault slip models (12 dip-slip and 7 strike-slip) using mainly dense strong-motion data and teleseismic data from Taiwan were utilized. Seven additional earthquakes (M >7) were included for further examination of scaling of large events. We determined the effective length and width for the scaling study was M0 ~ L2 and M0 ~ L3 for the events less than and larger than the seismic moment of 1020 Nm, respectively, regardless of the fault types, suggesting a nonself-similar scaling for small to moderate events and a self-similar scaling for large events. Although the events showed variation in stress drops, with the exception of three events with high stress drops, most of the events had stress drops of 10-100 bars. The observed bilinear relation is well explained by the derived magnitude-area equation of Shaw (2009) when we considered only events with stress drops of 10-100 bars and a seismogenic thickness of 35 km. The bilinear feature of the regressed magnitude-area scaling holds for ruptured areas up to about 1000 km2 for our seismogenic thickness of 35 km. For the events having rupture areas larger than that, the average slip becomes proportional to the rupture length. The distinct high stress drop events from blind faults in the western foothill of Taiwan yield local high peak ground accelerations (PGAs) when compared to the Next Generation Attenuation model. Regardless of the relative small magnitudes of these events, the high PGAs give the region higher seismic hazard potential and thus require special attention for seismic hazard mitigation.