At present, the classification prediction task based on news content or news headline has the problems of inaccurate classification and attention deviation. In this paper, a multi-model fusion attention network for news text classification (MFAN) is proposed to train news content and news titles in parallel. Firstly, the multi-head attention mechanism is used to obtain the category information of news content through a dynamic word vector, focusing on the semantic information that significantly influences the downstream classification task. Secondly, the semantic information of news headlines is obtained by using the improved version of the longshort-term memory network, and the attention is focused on the words that have a great influence on the final results, which improves the effectiveness of model classification. Finally, the classification fusion module fuses the probability scores of news text and news headlines in proportion to improve the accuracy of text classification. The experimental test on the TenthChina Software cup dataset shows that the F1 -Score index of the MFAN model reaches 97.789 %. The experimental results show that the MFAN model can effectively and accurately predict the classification of news texts.