The stochastic method for simulating strong ground motions from finite faults is applied to the records of the 1999 Chi-Chi, Taiwan, earthquake. The method involves discretization of the fault plane into smaller subfaults, each of which is assigned an x2 spectrum. The contributions from all subfaults are empirically atten- uated to the observation site and summed to produce the synthetic acceleration time history. The method is initially calibrated against the data recorded at 24 rock sites, located within 7-120 km from the mainshock hypocenter and providing a broad azimuthal coverage of the fault plane. The accuracy of the simulations is quantified through the model bias, defined as the logarithm of the ratio of the observed to simulated spectrum, averaged over all stations. The calibrated model for the Chi-Chi event has a near-zero average bias in reproducing the ground motions at rock sites in the frequency range from 0.1 to 20 Hz. An unusually low value is found for the radiation- strength factor s, controlling the high-frequency radiation level and directly related to the maximum slip velocity on the fault, compared with the mean value found for North American earthquakes. This result reflects the observed low peak ground ac- celerations of the Chi-Chi mainshock and, physically, its lower-than-usual slip ve- locities. The calibrated model is then used to simulate soil-site (site class D) records using the linear-response assumption. The simulated soil-site input motions are amplified by the weak-motion amplification functions, estimated by the spectral-ratio technique from available aftershock records. This analysis reveals an average reduction in strong-motion amplification to about 0.5-0.6 of that in weak motions, with an ac- celeration "threshold" for detectable nonlinearity near 200-300 cm/sec2. However, the derivation of soil-site specific weak-motion amplification was limited by the amount of aftershock data available; further improvement in the quantification of nonlinear soil response during the Chi-Chi earthquake may be possible with the release of additional aftershock datasets.