Google's WeatherNext 3 forecasts rain hourly, not every six hours

Google DeepMind and Google Research have released WeatherNext 3, an AI weather forecasting model that generates hourly predictions instead of the six-hour intervals typical of both traditional and prior AI systems. The finer time resolution lets the model track fast-moving weather, like a passing afternoon storm, with far more precision than earlier tools allowed.

On Operational WeatherBench, the industry benchmark for forecast accuracy, WeatherNext 3 outperformed rival models from Microsoft, Nvidia, and the European Center for Medium-Range Weather Forecasting. Google also says it beats traditional forecasts issued by the U.S. National Weather Service and ECMWF, the two institutions long considered the gold standard for physics-based prediction.

Rain forecasting saw the biggest jump: a 60% improvement over WeatherNext 2, the model's immediate predecessor. Google attributes part of that gain to sheer scale, WeatherNext 3 carries 2.4 times more parameters than the previous version, alongside architectural changes aimed specifically at precipitation.

Sharper resolution, satellite-native design

The model resolves key atmospheric variables at 5 km, a level of detail that matters for localized events like flash flooding or microbursts that coarser grids tend to smear across wide areas. Google is also billing WeatherNext 3 as the first high-resolution global model to ingest raw satellite observations directly, rather than relying on pre-processed data products.

That distinction matters competitively. Startup WindBorne has used raw-data satellite techniques since late 2025, and Google's move suggests the approach is becoming a contested edge in AI weather modeling rather than a niche technique.

WeatherNext 3 is also trained to align its output with specific ground weather stations, aiming to close the gap between global model predictions and the hyperlocal readings people actually check.

Where it shows up

Researchers will get access to WeatherNext 3 through Google's cloud platforms, continuing the company's pattern of open access for its DeepMind weather models. Consumer-facing deployment follows quickly: the model will power weather information across Search, Google Maps, and Gemini.

For a company that already dominates how billions of people check the forecast, folding a state-of-the-art model directly into Search and Maps turns an incremental accuracy gain into a change felt at planetary scale, one refreshed umbrella decision at a time.