Traffic flow prediction is beneficial to future intelligent traffic management and urban planning. However, existing studies have limitations in dealing with spatio-temporal dependencies and multimodal data fusion. It is difficult to comprehensively capture long-term dependencies and short-term spatial relationships, which limit prediction accuracy. In this paper, a multimodal traffic flow prediction model based on the Spatio-Temporal Mixed Attention network (ST-MANet) is proposed to address these issues in urban traffic flow prediction. The model's global attention is used to capture long-term trends and periodic patterns in historical data, while local attention is employed to capture short-term spatial relationships among adjacent regions or nodes. Moreover, the model fuses multimodal data, such as weather and holiday information, and introduces an adaptive attention scaling module to dynamically adjust the weights of global and local information, thereby enhancing the model's robustness and adaptability. Experimental results on the real traffic datasets PeMS04 and PeMS08 show that ST-MANet outperforms traditional methods and other mainstream models in terms of prediction precision and robustness.