FuXi-ENS is a machine learning model for medium-range ensemble weather forecasting. It outperforms the European Centre for Medium-Range Weather Forecasts (ECMWF) ensemble in 98.1% of 360 variable and forecast lead time combinations, based on the continuous ranked probability score (CRPS). The model generates 6-hourly forecasts up to 15 days at a spatial resolution of 0.25°, incorporating 5 atmospheric variables at 13 pressure levels and 13 surface variables. It uniquely combines CRPS and KL divergence loss, which is more effective than traditional L1 loss in ensemble forecasting. FuXi-ENS introduces perturbations at both initial conditions and each forecast step, enhancing forecast reliability. It is trained on 17 years of ECMWF ERA5 reanalysis data and demonstrates superior performance in deterministic and probabilistic metrics, including ensemble mean forecasts, extreme weather events, and tropical cyclone track forecasts. The model's computational efficiency allows it to generate 15-day forecasts in approximately 10 seconds per member on an Nvidia A100 GPU. FuXi-ENS also shows improved performance in predicting the 2018 Northeast Asia heatwave. The model's framework has potential for applications in other critical forecast scenarios requiring ensemble predictions, such as subseasonal-to-seasonal forecasts, seasonal forecasts, and wave forecasts. The study highlights the potential of machine learning in improving ensemble weather forecasting, particularly for extreme weather events where uncertainty quantification is essential.FuXi-ENS is a machine learning model for medium-range ensemble weather forecasting. It outperforms the European Centre for Medium-Range Weather Forecasts (ECMWF) ensemble in 98.1% of 360 variable and forecast lead time combinations, based on the continuous ranked probability score (CRPS). The model generates 6-hourly forecasts up to 15 days at a spatial resolution of 0.25°, incorporating 5 atmospheric variables at 13 pressure levels and 13 surface variables. It uniquely combines CRPS and KL divergence loss, which is more effective than traditional L1 loss in ensemble forecasting. FuXi-ENS introduces perturbations at both initial conditions and each forecast step, enhancing forecast reliability. It is trained on 17 years of ECMWF ERA5 reanalysis data and demonstrates superior performance in deterministic and probabilistic metrics, including ensemble mean forecasts, extreme weather events, and tropical cyclone track forecasts. The model's computational efficiency allows it to generate 15-day forecasts in approximately 10 seconds per member on an Nvidia A100 GPU. FuXi-ENS also shows improved performance in predicting the 2018 Northeast Asia heatwave. The model's framework has potential for applications in other critical forecast scenarios requiring ensemble predictions, such as subseasonal-to-seasonal forecasts, seasonal forecasts, and wave forecasts. The study highlights the potential of machine learning in improving ensemble weather forecasting, particularly for extreme weather events where uncertainty quantification is essential.