Google DeepMind and Google Research announced WeatherNext 3, a new artificial intelligence model designed to produce more frequent and finer-grained weather forecasts. Google says the model will begin feeding core variables into Search, Google Maps and Gemini, and will be available to users and researchers on Google’s cloud platforms.
Model design and capabilities
WeatherNext 3 is larger than its predecessor, with 2.4 times more parameters, and its designers adjusted decoder targets to produce more directly useful outputs. The team reports the model can predict key variables at about 5 km resolution and now produces hourly forecasts instead of the conventional six-hour cadence. Evaluations on rain performance are reported as 60% improved over WeatherNext 2.
Developers also trained the model to target forecasts for specific weather stations, allowing evaluations against ground-truth measurements. “Adding a capability where this model is now also predicting, say, what Denver’s airport’s weather station is going to measure on an hourly basis, just connects that forecasting task closer to the core,” said Daniel Rothenberg, an atmospheric scientist at Brightband.
Performance, data and comparisons
Google reports WeatherNext 3 was the most accurate among leading contenders on Operational WeatherBench, a benchmark built by Brightband that assesses metrics such as temperature, windspeed and humidity. According to the announcement, the model outperformed other deep-learning models from Google, Microsoft, Nvidia and ECMWF, and exceeded traditional forecasts from the U.S. National Weather Service and ECMWF.
The model’s increased update frequency is enabled by ingesting satellite observations collected hourly. Google describes WeatherNext 3 as the “first” AI model to directly incorporate raw observations for a high-resolution global forecast; the startup WindBorne says WeatherMesh 6 has used raw observations from balloons and other sources since late 2025. Google noted its forecasts offer higher resolution worldwide, while acknowledging both approaches still rely on national weather datasets for full performance.
Researchers highlighted broader impacts cited by public figures and specialists: Ferran Alet, a staff research scientist manager at DeepMind, framed machine learning as a way to learn patterns from large datasets; Bill Gates has pointed to AI-powered forecasts as a benefit for crop yields in developing countries; Alet added that higher-resolution forecasts of wind, rain and clouds may aid renewable energy projects.
Google says WeatherNext 3 will be integrated into select Google products and made available on its cloud platforms for further use and research.
Original source: TechCrunch AI