In order to improve the generalization ability of the twin support vector machine(TWSVM), a novel twin large margin distribution machine(TLDM) which increases the impact of the margin distribution on the training model is proposed. Theoretical studies show that margin distribution has important influence on the generalization performance of the model. The proposed approach based on the standard twin support vector machine adds the affection of margin distribution to the optimization objective function. The margin distribution is characterized by first order and second order statistics. The experimental results based on benchmark data sets show that the proposed approach has better classification accuracy than other three algorithms including SVM, TWSVM and TBSVM.