This article evaluates the long‐term probabilities of future earthquakes in the eastern Tibetan plateau, based on adaptively smoothed seismicity. The method is modified from Helmstetter et al. (2007) to allow for spatial and temporal variations of the completeness magnitude (), and to introduce spatially variable ‐value. By adjusting the algorithm of calculating kernel bandwidth for input events, earthquake data from varying intervals with unequal observation periods can be included into our model. Furthermore, a set of tests recommended by the Collaboratory for the Study of Earthquake Predictability have been carried out to test different options in the model application. Results show that our method can accurately forecast the number of observed events. Models with longer time duration of input catalog perform better resulting from the increase of input earthquakes. Forecasting effectiveness decreases with the target magnitude and is better for the Gaussian kernel compared with the Power‐law kernel. The approach selecting input events with the initial time of three intervals has only slightly better performance compared with simply using an 3.5, as only a very small number of earthquakes have been added by employing multiple intervals, resulting from the limit of time span and low quality of the input catalog. Nevertheless, our method contributes to improve the forecast effect with respect to spatially uniform models, which use input events since 1985 with a single ‐value of 0.862 for the study region. Detailed analyses demonstrate that using spatially variable ‐value or choosing input earthquakes with multiple intervals both contributes to model’s performance, and the effect of the former is more obvious in this study.