Stochastic blockmodel (SBM) has become a research focus in the domains of machine learning, network oriented data mining and social network analysis since it can effectively model networks without prior knowledge about their structures. It is a major challenge to develop a fast learning algorithm for stochastic blockmodel that has the capability of effective model selection for large-scale network. This paper presents a refined stochastic blockmodel, named RSBM, and its fast parallel learning method named RFLA. The learning method combines MML criteria with CEMM algorithm to achieve parallel execution in evaluating the model and estimating parameters. This strategy can significantly reduce time complexity of learning process. The accuracy and speed of the learning method are validated against both artificial networks and real networks, and the method is also compared with current representative SBM learning algorithms. The experimental results show that the proposed algorithm is able to greatly improve the efficiency without degenerating the precision of learning process, which indicates it achieves the best tradeoff between accuracy and speed. Furthermore, the proposed model and algorithm demonstrate the best generalization ability in terms of link prediction.