Brain tumor are uncontrollable and abnormal cells in the brain. The incidence and mortality of brain tumors are very high. Among them, gliomas are the most common primary malignant tumors with different degrees of invasion. The segmentation of brain tumors is a prerequisite for disease diagnosis, surgical planning and prognosis. According to the characteristics of brain tumor data, we designed a multi-model fusion brain tumor automatic segmentation algorithm based on attention mechanism [1]. Our network architecture is slightly modified based on 3D U-Net [2]. At the same time, the attention mechanism was added to the 3D U-Net model. According to the patch size and attention mechanism in the training process, four independent networks are designed. Here, we use 64 × 64 × 64 and 128 × 128 × 128 patch sizes to train different sub-networks. Finally, the results of the four models in the label layer are combined to get the final segmentation results. This multi model fusion method can effectively improve the robustness of the algorithm. At the same time, the attention method can improve the feature extraction ability of the network and improve the segmentation accuracy. Our experimental study on the newly released brats data set (brats 2019) shows that our method accurately describes brain tumors. © Springer Nature Switzerland AG 2020.