Google DeepMind Launches WeatherNext 3 with Hourly Forecasts and 5km Resolution

Google DeepMind and Google Research released WeatherNext 3, a forecasting model that updates hourly and offers a five-kilometer spatial resolution. [1][2] The system uses live geostationary satellite mosaics and atmospheric data to execute 24 forecast cycles per day. [1][2]

Major forecast runs occur four times daily to project 15 days ahead, while intermediate hourly runs produce 48-hour outlooks. [2] Google claims the model's precipitation predictions are up to 50 percent more accurate at lead times of 24 hours or longer. [2]

WeatherNext 3 is a Functional Generative Network mesh transformer. [1] The model trains directly on raw weather station measurements and satellite mosaics rather than numerical weather prediction reanalysis alone. [1] Under the hood, a 64-member ensemble runs in parallel to give probabilities of different weather scenarios. [2]

The forecasts are rolling out to Google Search, Google Maps, the Gemini mobile app, and enterprise cloud platforms via an allowlist. [1][2] However, the model weights remain closed-source and custom on-demand inference currently still relies on WeatherNext 2. [1]

By moving from the 25-kilometer, six-hour update cycle of WeatherNext 2 to a 5-kilometer hourly refresh rate, Google is tackling the localized timing challenges that traditionally plague global weather simulations. Using raw station measurements for training bypasses the geographical smoothing found in standard numerical models, better reflecting actual terrain-level variations. Independent evaluators like Brightband already consider it the most accurate global weather model. What the sources don't address: How much computational overhead is required to run the 64-member ensemble continuously compared to traditional physics-based models.