Working memory is critical to cognition. Recently, a delay estimation experiment for visual working memory has shown an interesting phenomenon that when given color stimuli in a uniform distribution, the distribution of color report is multimodal. The error between report and stimuli color follows a Gaussian like distribution with middle width. Previous model work can explain the heterogeneity of report but the error. We propose a spiking network model with heterogeneous connection and short-term plasticity. It can reproduce the distribution of report and error qualitatively.