To improve the accuracy of offline handwritten Chinese character recognition (offline HCCR), combined with self-attention, this paper proposes a collaborative multi-model approach for offline HCCR. Most existing offline HCCR models use different network structures to obtain different features, which may lead to different results in terms of accuracy, and the results that errors occurred may also be different. Utilizing this feature, combined with self-attention, we propose our method hoping to improve their accuracy. In this paper, five models, including AlexNet, VGG16, GoogLeNet, ResNet34 and ResNet50, are selected and modified as base models for offline HCCR, and the outputs of the adopted 2, 3, 4, and 5 models are corrected by the self-attention fusion module. Our methods are tested on the evaluation dataset of the ICDAR 2013 Chinese Handwriting Recognition Contest. Our HCCR results obtained are at least 0.485, 0.786, 0.981 and 1.065 percentage points higher than the highest accuracy of all 2, 3, 4 and 5 base models, respectively. The experiments show that our method is effective in improving the accuracy of offline HCCR.