FengWu-GHR is a groundbreaking machine learning-based global numerical weather prediction model that operates at a kilometer-scale resolution, making it the first AI model to achieve this level of detail. The model addresses the challenges of high-resolution forecasting by leveraging a pretrained low-resolution model and incorporating transfer learning techniques. Key features of FengWu-GHR include: 1. **High Resolution**: Operating at a 0.09° horizontal resolution, FengWu-GHR provides more accurate and detailed weather forecasts compared to existing models. 2. **Large Parameter Count**: The model has over 4 billion trainable parameters, making it the most parameter-rich forecasting model in atmospheric science. 3. **General Framework for Downscaling**: The model's design allows for the advancement of many ML-based weather forecast models from low to high resolution with minimal computational cost. The model's performance is evaluated using various metrics, including RMSE, ACC, bias, and activity, and compared against the operational version of the Integrated Forecasting System (IFS-HRES). Results show that FengWu-GHR outperforms IFS-HRES in 91.3% of the target variables, demonstrating superior skill in weather forecasting, especially in station observations and extreme event predictions. The model also shows reduced bias drift and improved stability over long lead times, making it a promising tool for operational weather forecasting and climate research.FengWu-GHR is a groundbreaking machine learning-based global numerical weather prediction model that operates at a kilometer-scale resolution, making it the first AI model to achieve this level of detail. The model addresses the challenges of high-resolution forecasting by leveraging a pretrained low-resolution model and incorporating transfer learning techniques. Key features of FengWu-GHR include: 1. **High Resolution**: Operating at a 0.09° horizontal resolution, FengWu-GHR provides more accurate and detailed weather forecasts compared to existing models. 2. **Large Parameter Count**: The model has over 4 billion trainable parameters, making it the most parameter-rich forecasting model in atmospheric science. 3. **General Framework for Downscaling**: The model's design allows for the advancement of many ML-based weather forecast models from low to high resolution with minimal computational cost. The model's performance is evaluated using various metrics, including RMSE, ACC, bias, and activity, and compared against the operational version of the Integrated Forecasting System (IFS-HRES). Results show that FengWu-GHR outperforms IFS-HRES in 91.3% of the target variables, demonstrating superior skill in weather forecasting, especially in station observations and extreme event predictions. The model also shows reduced bias drift and improved stability over long lead times, making it a promising tool for operational weather forecasting and climate research.