Abstract: Loess seismic landslides are common and highly hazardous earthquake-induced geological disasters on the Loess Plateau in Northwest China. Establishing a critical displacement criterion for instability that incorporates seismic dynamic response is essential for identifying co-seismic landslides and for early warning of earthquake-induced landslides. This study adopts a typical loess–mudstone slope as the prototype and performs shaking-table model tests. Recorded waveforms, including the Wolong Station record from the 2008 Wenchuan earthquake and the El Centro Station record from the 1940 Imperial Valley earthquake (USA), are used as input motions. The amplification of peak ground acceleration (PGA) and Arias intensity in the test model (geometric similarity ratio 1: 10) is examined under horizontal seismic excitations with acceleration amplitudes from 0.05 g to 0.8 g to elucidate the dynamic response of the model slope. By integrating macroscopic slope deformation with micro-monitoring data of acceleration and gyroscopic displacement, the seismic deformation and failure processes are analyzed, and critical displacement criteria are established for seismic cracking of the loess slope mass and for instability-induced sliding. The results indicate that, as the input seismic PGA amplitude increases from low (0.05 g–0.2 g) to moderate (0.2 g–0.4 g) and high levels (0.4 g–0.6 g), the amplification factors of PGA and Arias intensity exhibit a nonlinear pattern characterized by an initial increase, followed by a decrease, and a subsequent increase. Overall, seismic amplification increases with slope height, and Arias-intensity amplification exceeds that of PGA. In the upper slope, the maximum Arias-intensity amplification factor reaches 2.5, whereas the maximum PGA amplification factor is 1.8. The amplification factor at the slope surface is slightly higher than within the slope interior but lower than near the potential slip surface. Within the slope, amplification in the bedrock is weaker than in the loess layer. Comparative analysis of the slope responses to the Wolong and El Centro records indicates that the input ground motion and the slope-site fundamental frequency jointly govern the seismic response; the amplification effect is greatest when the predominant period of the input motion closely matches the site’s fundamental period. Under seismic loading, loess landslides exhibit a cracking–sliding failure mode. In the model slope, microcracks initiate in the upper portion at a shaking-table input of 0.3 g; at 0.5 g, surface cracks evolve from fine, short fissures into deep, vertically penetrating cracks, corresponding to a critical displacement of 3.5 cm for the cracked slope mass. At 0.6 g seismic loading, crack coalescence occurs, and sliding is triggered along the loess–mudstone interface, corresponding to a critical displacement of 9.85 cm for landslide instability. These findings are expected to enhance the accuracy of co-seismic landslide identification and advance quantitative hazard assessment.