We modeled the source rupture process of the 20 September 1999 Chi-Chi, Taiwan, earthquake (Mw 7.6) using an integrated dataset of near-field strong ground motion and Global Positioning System observations. This large inland thrust earthquake produced a surface break more than 80 km long along the Chelungpu fault. The event triggered over 700 strong-motion instruments on the island of Taiwan. This large ground-motion dataset provides the best opportunity to date to understand earthquake source processes in a major earthquake. First, we develop a 3D curved fault model consistent with the surface rupture data and aftershock distribution. The subsurface fault plane has an average dip of 30° to the east, and measures 96 km along strike and 40 km down-dip. Next we applied a genetic algorithm to reconstruct the source rupture process. The inversion results indicate a complex slip distribution with a generally broad triangular shaped slip zone that points down-dip. Most of the displacement occurred over a 15-km depth range in the northern segment of the fault, with the maximum slip exceeding 20 m. The average slip is 3.8 m, and the average rupture velocity is 2.6 km/sec. The mainshock slip distribution clearly complements the aftershock location distribution. Slip rise times range from 8 to 6 sec over most of the primary slip area and decrease slightly as rupture propagates south to north. This relatively long rise time resulted in a low dynamic stress drop and low peak ground accelerations observed at surface stations. The rake angle rotates across the rupture plane from predominantly dip-slip in the southern segment to mostly left-lateral oblique slip in the north. We find that geometrical irregularities over the fault plane played an important role in controlling the temporal rupture behavior. At fault bends, rupture tends to decelerate, the rake angle rotates to a near dip-slip orientation, and the shear stress increases as rupture tries to overcome the bending barriers. The total seismic moment from our inversion is 2.9 × 1027 dyne cm, in good agreement with the Harvard and the U.S. Geological Survey moment estimation.