Large earthquakes, such as Wenchuan in 2008, 7.9, Sichuan, China, provide an opportunity for earthquake early warning (EEW), as many heavily shaken areas are far () from the epicenter and warning times could be sufficient () to take preventive action. On the other hand, earthquakes with magnitudes larger than 6.5 are challenging for EEW because source dimensions need to be defined to adequately estimate shaking. Finite‐fault rupture detector (FinDer) is an approach to identify fault rupture extents from real‐time seismic records. In this study, we playback local and regional onscale strong‐motion waveforms of the 2008 7.9 Wenchuan, 2013 6.6 Lushan, and 2017 6.5 Jiuzhaigou earthquakes to study the performance of FinDer for the current layout of the China Strong Motion Network. Overall, the FinDer line‐source models agree well with the observed spatial distribution of aftershocks and models determined from waveform inversion. However, because FinDer models are constructed to characterize seismic ground motions (as needed for EEW) instead of source parameters, the rupture length can be overestimated for events radiating high levels of high‐frequency motions. If the strong‐motion data used had been available in real time, 50%–80% of sites experiencing intensity modified Mercalli intensity IV–VII (light to very strong) and 30% experiencing VIII–IX (severe to violent) could have been issued a warning with 10 and 5 s, respectively, before the arrival of the S wave. We also show that loss estimates based on the FinDer line source are more accurate compared to point‐source models. For the Wenchuan earthquake, for example, they predict a four to six times larger number of fatalities and injured, which is consistent with official reports. These losses could be provided faster than if they were based on more complex inversion rupture models.