A probabilistic weather forecast is critical for decision-making, from the everyday to preparation for extreme events. Forecasts of the joint distribution of weather trajectories via spatio-temporally coherent ensembles have further importance: they provide a powerful tool for decision-making in complex and high-impact domains including energy system planning, transportation routing, flood forecasting and more. State-of-the-art ML forecast models for medium-range weather, however, are largely trained to produce deterministic forecasts which lose physical consistency at longer lead times. We introduce GenCast, a ML-based generative model for ensemble weather forecasting, trained from reanalysis data. It forecasts ensembles of trajectories for hundreds of weather variables, up to 15 days at 1 degree resolution globally, using under TC secs per ensemble member. We show that GenCast is more skilful than ENS, a top operational ensemble forecast, for TC% of TC verification targets, while maintaining good calibration and physically consistent power spectra.