Swift: an autoregressive consistency model for efficient weather forecasting

Diffusion models offer a physically grounded framework for probabilistic weather forecasting, but their typical reliance on slow, iterative solvers makes inference expensive, especially for forecasting at long-lead times with a high temporal resolution. To address this, we introduce , a single-step consistency model that, for the first time, enables autoregressive finetuning of a probability flow model with a continuous ranked probability score objective. This eliminates the need for multi-model ensembling or parameter perturbations. Results show that produces skillful 6-hourly forecasts that match the skill of our diffusion baseline with significantly lower sampling cost, and remains stable in 75-day autoregressive rollouts. Specifically, requires $39\times$ fewer neural function evaluations per autoregressive step than diffusion baselines, while achieving medium-range forecast skill competitive with the numerical-based, operational IFS ENS. These results demonstrate that consistency models can make probabilistic weather forecasting substantially more efficient while opening a practical path to post-training calibration.