Google DeepMind's WeatherNext 3 AI weather model brings 5km hourly forecasts to Search, Maps and Gemini

On September 3, scientists at Google DeepMind and Google Research released WeatherNext 3, a new artificial intelligence model for weather forecasting that Google says sees the changing atmosphere more clearly and predicts its behavior more often. The release is the latest wave of a sea change that deep learning has brought to meteorology, and it arrives with unusually concrete product plans.

Google says WeatherNext 3 will start feeding into the weather information users see in Search, Google Maps and Gemini, and will also be available to users and researchers on Google's cloud platforms. “This is going to be the first time that some of the core variables feed and power a lot of the Google products,” Samier Merchant, a Google senior staff engineer, told TechCrunch.

The model's accuracy has been validated by an independent comparison: WeatherNext 3 already proves to be the most accurate among leading contenders tested on Operational WeatherBench, a utility for comparing AI forecasts built by the startup Brightband. On metrics such as temperature, windspeed and humidity, it beats other deep-learning models from Google, Microsoft, NVIDIA and the European Centre for Medium-Range Weather Forecasts (ECMWF), as well as traditional forecasts from the US National Weather Service and the ECMWF.

For decades, most forecasts have come from government-owned supercomputers churning through mathematical equations that describe the physics of the atmosphere; these systems are remarkably accurate but expensive and comparatively slow. After the ECMWF released more than half a century of weather data in 2018, deep-learning researchers began training models that can make predictions far more quickly and with comparable accuracy.

Model-makers have since chipped away at the known weaknesses of AI forecasters: they tend to predict over wide areas of 15 to 25 square kilometers, they are not always good with rain, and they still depend on formatted datasets produced by government agencies. WeatherNext 3 was designed to take on all three challenges at once.

On key variables, researchers told TechCrunch, WeatherNext 3 can predict down to a resolution of 5 kilometers, its evaluations on rain are 60% improved over WeatherNext 2, and it can now produce hourly forecasts instead of the standard prediction every six hours. The model is 2.4 times larger in parameters than its predecessor, with decoder heads tailored to give more useful answers, and it was trained to target forecasts at specific weather stations so its output can be checked against ground-truth observations.

The higher forecast frequency comes from ingesting weather-satellite data collected in real time on an hourly basis. Google calls WeatherNext 3 the “first” AI model to directly incorporate raw observations for a high-resolution global forecast, but AI weather startup WindBorne says its WeatherMesh 6 model has been doing so since late 2025; Google counters that its own forecasts are higher resolution across the globe. Both models still rely on national weather datasets, so true end-to-end data assimilation remains future work.

The weather domain is a reminder that the transformer revolution matters well beyond chatbots. European and US weather agencies are already using AI models in their forecast products, and the speed and low cost of these systems promise to bring dependable forecasts to poorer regions where supercomputers and dense sensor networks are out of reach. DeepMind researchers also point to higher-resolution wind, rain and cloud forecasts as a way to make renewable energy projects more dependable — the payoff worth watching as WeatherNext 3 reaches Google's products.