Google DeepMind Launches WeatherNext 3, Advanced AI Weather Model

On September 5, 2026, Google DeepMind and Google Research introduced WeatherNext 3, an AI-driven weather forecasting system designed to generate hourly global forecasts using live geostationary satellite data and sparse weather station readings. This model represents a significant advancement in forecasting accuracy and is now being integrated into various Google services, including Search, Gemini, Maps, and Google Cloud. Weather has a profound impact on numerous aspects of daily life, influencing decisions from carrying an umbrella to managing supply chains and power grids during extreme weather events. By addressing the challenges of predicting fast-moving, hyper-local weather conditions, WeatherNext 3 aims to enhance the reliability of weather forecasts for users worldwide.

Historically, weather forecasting models have relied on traditional numerical weather prediction systems, which use physics-based simulations that are limited by a six-hour data lag. This delay can introduce inaccuracies for rapidly changing weather variables such as rainfall and surface temperature. WeatherNext 3 innovates by utilizing a continuously updating stream of live geostationary satellite data, enabling it to produce fresh forecasts every hour. Additionally, it directly incorporates readings from sparse weather stations, allowing it to capture sharp local variations in weather that older models often smooth over.

The new model's resolution is impressive, achieving surface variable resolutions of 5 kilometers, other surface variables at 10 kilometers, and atmospheric variables like wind speed at 25 kilometers—significantly more detailed than its predecessor, WeatherNext 2. Implementation of WeatherNext 3 involved collaboration among key stakeholders, including meteorologists, data scientists, and engineers at Google. The model's development focused on overcoming previous limitations in precipitation forecasting, which has historically posed challenges due to the complexities of small-scale cloud processes.

Google trained WeatherNext 3 using two high-quality precipitation datasets: NASA's satellite-based IMERG product and its radar-based global precipitation reanalysis. As a result, the model has shown substantial accuracy improvements in medium-range forecasts, with gains of up to 60% compared to satellite benchmarks, 30% against radar data, and 10% against ground rain-gauge measurements. The launch of WeatherNext 3 not only enhances forecasting capabilities but also provides significant benefits for clean energy planning. The model offers predictions that include turbine-height wind speeds and detailed estimates of cloud cover and solar radiation, which are crucial for grid operators and renewable energy developers.

This is especially important for regions in Latin America, Africa, and the Asia-Pacific where access to high-resolution forecasting has been limited due to the high computational resources required for traditional models. WeatherNext 3 aims to bridge this gap, providing more accurate forecasts to underserved areas. Looking ahead, WeatherNext 3 is set to transform the landscape of weather forecasting. It began powering features in Google Search, Gemini, Google Maps, and the Google Maps Platform Weather API immediately following its announcement.

Users can expect enhanced accuracy in precipitation forecasts, particularly in regions where traditional forecasting has been less effective. The underlying hourly, high-resolution data is also accessible to developers and researchers through BigQuery, Earth Engine, and Google Cloud Storage. While WeatherNext 3 offers advanced forecasting capabilities, Google emphasizes that it does not replace official sources for severe weather warnings or public safety advisories, which should continue to come from local meteorological agencies and national weather services.