A new forecast every hour
Google DeepMind and Google Research introduced WeatherNext 3 on September 3, 2026. The model generates global probabilistic forecasts every hour, incorporates low-latency geostationary satellite imagery, and increases the detail of several surface predictions compared with WeatherNext 2. Google began integrating its output into Search, Maps, Gemini, Earth Engine, Cloud, and developer products at launch.
Earlier AI weather models largely learned from reanalysis or analysis fields produced by conventional numerical weather-prediction systems. Those datasets combine observations with physics-based models, but they arrive on fixed cycles and can carry the biases of the assimilation system. WeatherNext 3 adds raw satellite observations directly to the forecasting pipeline, allowing a new initialization each hour and providing a more immediate view of changing clouds and atmospheric structure.
Resolution and new target variables
The model produces some station-targeted temperature and humidity information at roughly five-kilometre resolution, other surface fields around ten kilometres, and upper-atmosphere fields at coarser resolution. Its paper describes hourly time steps and a flexible generative architecture that predicts dense global grids, station values, precipitation estimates, and cyclone information.
WeatherNext 3 also targets variables useful for renewable energy, including solar radiation, cloud cover, and wind near turbine height. More frequent ensemble forecasts can help an operator estimate a range of possible generation rather than plan around one deterministic line.
Google reported substantial precipitation-score improvements against satellite, radar, and gauge targets, with the size varying by dataset and forecast lead time. It also cited independent live evaluations by Brightband. Those results are promising, but the launch figures cover selected metrics and periods. Weather skill varies by region, season, variable, and extreme event; users need local verification rather than one global average.
Why probabilistic forecasting matters
Weather is chaotic. Small uncertainty in today’s atmosphere expands with forecast time, so an honest system should describe multiple plausible futures. WeatherNext 3 is an ensemble model, producing many scenarios that can be summarized as probabilities, ranges, or risk thresholds.
That is particularly useful for decisions such as grid balancing, logistics, farming, and disaster preparation. A 30 percent chance of a damaging event can justify action even when the most likely single forecast does not show it. The quality of calibration matters as much as the sharpness of an image: an event assigned a 30 percent probability should occur about that often across comparable cases.
The model’s speed can make large ensembles and frequent updates practical, but AI forecasts do not abolish physics-based prediction. Operational centers combine observations, multiple models, forecaster expertise, and knowledge of local terrain and infrastructure. Physics-based systems also remain essential for research, data assimilation, and comparison when a learned model encounters unusual conditions.
Limits, data gaps, and public warnings
Satellite coverage and quality differ across the planet, and geostationary sensors provide a less direct view near the poles. Weather stations are unevenly distributed. Training on historical observations can underrepresent unprecedented extremes or changes in observing systems. Fine grid spacing also does not mean the model explicitly resolves every cloud or neighbourhood-scale hazard.
A forecast integrated into a consumer product is not an official warning. Google’s own announcement directs people to national meteorological agencies and local services for severe-weather alerts. Those agencies combine forecasts with radar, emergency thresholds, communication systems, and legal responsibility.
The system is also a Google-developed model deployed through Google infrastructure. Researchers received access to forecast data and a paper, but full independent reproduction requires substantial data and compute. Evaluation should include failure cases, not only average accuracy, because rare missed extremes can dominate public impact.
Why it matters
WeatherNext 3 moved AI forecasting closer to an end-to-end operational system that learns from recent observations, updates hourly, produces ensembles, and serves high-resolution outputs through widely used products. It showed AI contributing beyond faster emulation of a traditional forecast toward a different pipeline for observation, prediction, and post-processing.
The advance is most valuable as part of a forecasting ecosystem. Better models can give experts more timely evidence, but accountable warnings still require human institutions, multiple data sources, and clear communication of uncertainty.