WeatherNext 3: Google DeepMind’s hourly, satellite-powered global weather model

# What WeatherNext 3 is and why it matters

# How the model works in plain terms The system ingests real-time satellite imagery plus traditional historical data into a single Functional Generative Network (FGN) mesh transformer. From that input it produces several forecast products: dense gridded fields of meteorological variables, discrete cyclone tracks, and native station-level predictions. The end-to-end design maintains broad atmospheric consistency while resolving local topography and surface effects.

# Resolution and refresh cadence WeatherNext 3 produces forecasts at multiple native spatial scales:

• 5-kilometer resolution for key surface variables like 2-meter temperature and moisture.
• 10-kilometer resolution for other surface variables.
• 25-kilometer resolution for atmospheric variables such as wind speed.

The model issues updates every hour. Compared with WeatherNext 2—which ran on a 25-kilometer grid with six-hour increments—WeatherNext 3 is roughly five times sharper and refreshes much more frequently.

# Outputs and technical features

An independent live evaluation by Brightband is cited in the announcement as part of the model's validation.

# Where you'll see it and how to access the forecasts WeatherNext 3 is integrated across Google products: Search, Gemini, Maps, Google Maps Platform, and Google Cloud. That means end users will receive the enhanced forecasts in consumer-facing Google apps, while businesses and developers can consume the model's outputs through Google Cloud and the Maps Platform.

# Practical use cases

• Daily planning: more reliable hour-by-hour precipitation and temperature predictions for individuals.
• Renewable energy: higher-resolution wind-at-turbine-height and solar irradiation inputs for grid operators and plant operators.
• Agriculture and supply chains: finer-grained moisture and temperature forecasts for short-term decisions.
• Emergency and storm tracking: discrete cyclone outputs and higher-frequency updates for rapidly evolving events.

# What the announcement emphasizes and what it does not claim The announcement emphasizes a move to live satellite data and hourly refreshes to resolve local and rapidly changing weather. The release highlights higher native spatial resolution, multiple output types, and product integration across Google services. The post references independent live evaluations (Brightband) without reproducing those evaluation details in full.

# Takeaway WeatherNext 3 focuses on timelier, higher-resolution forecasts by training on continuous satellite observations and producing native, hourly outputs at multiple spatial scales. The model's integration into Google's consumer products and Cloud tools aims to make those higher-frequency, higher-resolution forecasts broadly available to both users and developers.