A global validation dataset of 116 seismic events and 20,977 associated Pn and P arrivals is used to assess travel‐time prediction and event location accuracy for the global‐scale, 3D, P‐wave velocity model called LLNL‐G3Dv3 (Simmons et al., 2012). Strong regional trends that are observed for ak135 travel‐time residuals are largely removed when LLNL‐G3Dv3 is used for prediction. The 25th–75th quantile spread of travel‐time residuals is reduced by 30%–40% at teleseismic distances, and the spread is reduced by ∼60% at regional distances (<16°). Epicenter error decreases when more data are used to constrain event locations until more than ∼40 arrivals times are used. At which point, epicenter error reduction tends to plateau. Median epicenter errors for the ak135 and LLNL‐G3Dv3 models plateau at ∼8.0 and ∼5.5 km, respectively, for teleseismic P datasets. Median epicenter errors for the ak135 and LLNL‐G3Dv3 models plateau at ∼12.0 and ∼4.0 km, respectively, for regional Pn datasets. We demonstrate that spatially correlated travel‐time residual errors for the ak135 model lead to increasing epicenter error when ∼40 to ∼100 Pn arrivals are used to constrain the location. The effect of correlated error is mitigated by LLNL‐G3Dv3, for which epicenter error steadily decreases to ∼4 km when 100 Pn arrivals are used. The median area of 0.95 epicenter probability bounds for ak135 and LLNL‐G3Dv3 are 1811 and 758 km2, respectively. The ak135 ellipses are inflated to achieve the desired rate of true events occurring inside the probability region, whereas LLNL‐G3Dv3 error ellipses based on empirical residual distributions cover the true location at the expected rate because location bias is minimal.