Inferring dynamic complex network through a small set of samples is a challenging problem in the field of biological network, social network and transportation network, which can help improve understanding of complex network systems. In this letter, a new Hybrid Model based Latent Variables Sampling algorithm is presented to address the problems of high computation complexity and low accuracy faced by traditional approaches. Experimental results on simulated and real data sets show that the presented method possesses better reasoning performance and significantly improves the precision and efficiency of network inference especially when compared with the other three approaches. Under different dimensions, HM-LVS still has higher accuracy (average 80%) and can effectively reverse engineering dynamic complex networks from time series data.