Google has released WeatherNext 3, an advanced artificial intelligence weather forecasting model designed to deliver faster and sharper predictions by ingesting live satellite observations, according to company announcements. The new deep-learning model produces global forecasts five times sharper than previous iterations and integrates directly into consumer applications including Google Search, Maps, and Gemini.
How WeatherNext 3 Outperforms Traditional Forecasts
Weather forecasting has historically relied on government-owned supercomputers running complex physics equations. While accurate, those traditional systems require significant computing power and time. According to Google Research senior staff engineer Samier Merchant, deep learning models bypass these delays by recognizing patterns in massive datasets.
WeatherNext 3 improves upon its predecessor, WeatherNext 2, by scaling up its parameter count by 2.4 times and tuning its architecture to predict specific atmospheric variables down to a 5-kilometer resolution, compared to the older 25-kilometer grid. According to Google DeepMind staff research scientist manager Ferran Alet, the model directly tackles noisy physics and incomplete information by learning empirical patterns from historical and real-time data.
| Metric | WeatherNext 2 | WeatherNext 3 |
|---|---|---|
| Resolution | 25-kilometer grid | Up to 5-kilometer resolution |
| Update Frequency | Every 6 hours | Hourly forecasts |
| Precipitation Accuracy | Standard baseline | Up to 50% more accurate 24 hours in advance |
Integration Across Global Products and Renewable Energy
Google is rolling out the capabilities of WeatherNext 3 across its core ecosystem. Users will encounter the updated weather data directly within Google Search, Google Maps, and the Gemini AI assistant. Furthermore, the model is available to external researchers and enterprise users on Google’s cloud platforms.
Beyond daily rain and temperature checks, the model targets renewable energy planning. Ferran Alet noted to The Verge that the architecture evaluates wind speeds at 100 meters—roughly the hub height of a modern wind turbine—to help optimize green energy generation for data centers and the broader power grid.
Addressing Regional Gaps in Global Forecasting
A key hurdle for artificial intelligence meteorology has been its reliance on heavily formatted data originating primarily from the United States and Europe. WeatherNext 3 incorporates raw, real-time hourly satellite observations to close these gaps. Atmospheric scientist Daniel Rothenberg of Brightband noted that predicting metrics down to specific local stations, such as Denver’s airport weather station, bridges the gap between broad AI models and ground-truth measurements.
Independent benchmarking reinforces these advances. According to Operational WeatherBench, a comparison utility built by Brightband, WeatherNext 3 outperforms competing deep-learning frameworks from Microsoft, Nvidia, and the European Center for Medium-Range Weather Forecasting, alongside traditional government models from the US National Weather Service.
Worth a look