Swift: An Autoregressive Consistency Model for Efficient Weather Forecasting

Diffusion models offer a physically grounded framework for probabilistic weather forecasting but suffer from slow, iterative inference that make adapting to domain-specific objectives, such as the continuous ranked probability score (CRPS), impractical. We address these limitations by pretraining a single‐step consistency model that generates diverse ensembles in one forward pass, removing the need for multi-model ensembling or parameter perturbations. Crucially, we demonstrate for the first time the end-to-end finetuning of a probability flow generative model on a multi-step CRPS loss. Our results yield fast and calibrated autoregressive forecasts stable to 75 days, competitive with the state‐of‐the‐art at a fraction of the compute.