Urban traffic prediction is crucial for improving navigation and traffic management efficiency in increasingly congested cities worldwide. While data-driven models have shown strong performance in spatio-temporal forecasting, their effectiveness is often limited by a reliance on single-source data. External factors, such as weather conditions and local events, offer significant potential for enhancing prediction accuracy. This paper introduces a novel Spatio-Temporal Multimodal Mixture-of-Experts (STMMoE) model, which effectively integrates attribute and structure information inherent in traffic data to learn complex spatio-temporal correlations. By fusing multimodal historical data, including traffic flow, weather, and event information, our model can process arbitrary input and output lengths, making it applicable across large urban areas. Our extensive experiments with real-world datasets from London demonstrate that our proposed STMMoE model achieves outstanding performance. It significantly outperforms traditional baselines, showcasing its capability for accurate and reliable long-term traffic forecasting, which is essential for developing advanced traffic management systems. We release the datasets at: https://github.com/TPUrbanNav/LM_Dataset.git