This article addresses the forward simulation of regional and local strong motions from the 1923 Kanto, Japan, earthquake (Ms 8.2), using two variable-slip rupture models: the Wald and Somerville (1995) slip model (WS slip model) derived from geodetic and teleseismic data and a resampled version of the Takeo and Kanamori (1992) slip model (TK slip model) that was obtained from the forward simulation of the Ewing seismoscope record at Hongo (HNG, epicentral distance R = 60 km). Green's functions are calculated using flat-layered velocity models for specific source-receiver paths, which were developed in our first article (Sato et al., 1998). For regional stations Sendai (R = 350 km) and Gifu (R = 220 km), the WS slip model provides a much better match to the waveform data than the TK slip model. For local station HNG, the first 40 sec of the N77°E Imamura seismogram is successfully reproduced by both of the slip models, although the arrivals are delayed in the TK slip model. The large-amplitude long-duration, long-period (13 sec) later phases in the southwest-component Ewing seismogram are not reproduced by either of the slip models. Our results suggest that the WS slip model gives a better representation of the overall rupture process of the 1923 event than the TK slip model does. Near-fault long-period ground motions calculated at several stations using the WS slip model suggest that the motions at HNG were not the largest in the Tokyo metropolitan area during the 1923 event. In addition, we estimate that the ground motions near the southern margin of the fault plane were significantly larger than the recorded near-fault motions of recent magnitude 7 earthquakes, such as Northridge and Kobe, for periods longer than several seconds. This suggests that design codes based on the experience of these recent events may not adequately describe the long-period response expected during a magnitude 8 earthquake.