WeatherNext 2 release: DeepMind's advanced weather model

DeepMind released WeatherNext 2 this week, a second-generation global AI model designed to deliver higher-resolution, more accurate short- to medium-range forecasts while reducing compute per forecast. The model produces gridded global outputs at finer spatial scales and includes an explicit probabilistic ensemble stage to estimate uncertainty.

WeatherNext 2 follows the company’s earlier weather projects and targets operational forecasting use cases where speed and spatial detail matter. DeepMind highlights improvements in forecast skill at lead times from hours to several days, claims a reduction in inference cost compared with its prior version, and emphasizes the model’s ability to fill gaps where traditional models struggle to run high-resolution global ensembles.

How WeatherNext 2 works

The architecture combines dense spatio-temporal neural components with preprocessing and postprocessing layers that convert observational and reanalysis inputs into operational forecast grids. Inputs include satellite radiances, radar composites, surface observations, and historical reanalysis fields. Those inputs pass through quality control and regridding stages before entering the neural core.

The neural core operates on a global grid and is trained to predict the evolution of multiple atmospheric fields simultaneously. DeepMind describes a stacked design that separates short-range dynamics from medium-range patterns, and a downstream ensemble module produces probabilistic outputs by perturbing inputs and model states. Final postprocessing applies bias correction and physical consistency checks before producing user-facing forecast maps and probabilistic products.

Training used a mix of historical reanalysis and modern observations, with data augmentation and physically informed loss functions to keep outputs consistent with known atmospheric relationships. DeepMind also says it optimized the model for inference efficiency, enabling higher spatial resolution without a proportional rise in compute compared with its previous release.

Validation, benchmarks and availability

DeepMind published benchmark comparisons showing WeatherNext 2 outperforming its predecessor on several standardized skill metrics and improving precipitation and near-surface temperature forecasts at short leads. The company also highlights reduced runtime for equivalent-resolution forecasts, positioning the model as a candidate for operational ensemble forecasting where compute cost is a bottleneck.

The release includes technical documentation and examples of regional and global forecast products. DeepMind indicates collaborators and academic partners will have access for evaluation, and it expects operational agencies and private weather services to test integration with existing ingest and dissemination systems. Licensing terms and deployment options vary by partner; DeepMind says the model is intended for research and operational trials rather than a consumer product.

Independent verification remains important. The model’s performance will depend on local conditions, the quality of ingest streams, and integration with local postprocessing. Operational centers will likely run parallel comparisons before replacing or augmenting existing suites.

Why it matters

WeatherNext 2 signals continued interest from large AI labs in building practical scientific models that target operational constraints such as compute cost and ensemble forecasting. If the claimed gains in accuracy and efficiency hold up under independent evaluation, the model could speed higher-resolution forecasting in regions that currently lack the compute budgets for large numerical model ensembles. That would affect meteorological services, emergency planners, and industries that rely on timely high-resolution weather information.