The paper introduces *XiHe*, a data-driven 1/12° resolution global ocean eddy-resolving forecasting model. *XiHe* is designed to address the limitations of traditional physics-driven numerical models, which are computationally expensive and slow, and often struggle with improving forecasting accuracy due to their reliance on human understanding of physical laws. *XiHe* leverages a hierarchical transformer framework, incorporating a land-ocean mask mechanism and an ocean-specific block to focus on global ocean dynamics and capture both local and global ocean information. Extensive experiments using satellite and in situ observations, as well as the IV-TT Class 4 evaluation framework, demonstrate that *XiHe* outperforms leading operational numerical GOFSSs in terms of forecast accuracy, especially for ocean current forecasting up to 60 days. *XiHe* also shows superior performance in forecasting large-scale ocean circulation and mesoscale eddies, with a forecast speed that is thousands of times faster than traditional numerical models.The paper introduces *XiHe*, a data-driven 1/12° resolution global ocean eddy-resolving forecasting model. *XiHe* is designed to address the limitations of traditional physics-driven numerical models, which are computationally expensive and slow, and often struggle with improving forecasting accuracy due to their reliance on human understanding of physical laws. *XiHe* leverages a hierarchical transformer framework, incorporating a land-ocean mask mechanism and an ocean-specific block to focus on global ocean dynamics and capture both local and global ocean information. Extensive experiments using satellite and in situ observations, as well as the IV-TT Class 4 evaluation framework, demonstrate that *XiHe* outperforms leading operational numerical GOFSSs in terms of forecast accuracy, especially for ocean current forecasting up to 60 days. *XiHe* also shows superior performance in forecasting large-scale ocean circulation and mesoscale eddies, with a forecast speed that is thousands of times faster than traditional numerical models.