Google's new AI weather model forecasts down to 5km

Google DeepMind and Google Research released a new AI weather-forecasting model called WeatherNext 3, which the company says sees the atmosphere more clearly and predicts its behavior more often than before. Google says the model will start feeding weather information into 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,” Google senior staff engineer Samier Merchant told TechCrunch.

WeatherNext 3 has already proven the most accurate model among leading contenders on Operational WeatherBench, a tool built by startup Brightband to compare AI forecasts on metrics like temperature, windspeed, and humidity — beating out deep-learning models built by Google, Microsoft, Nvidia, and the European Center for Medium-Range Weather Forecasting (ECMWF), as well as traditional forecasts from the U.S. National Weather Service and the ECMWF.

The new model addresses three key weaknesses that have limited AI weather forecasting so far: it can predict key variables down to a resolution of 5 kilometers, versus the wider 15-to-25-kilometer area typical of earlier models; its rain evaluations are 60% improved over WeatherNext 2; and it produces hourly forecasts instead of the standard prediction every six hours. WeatherNext 3 is also 2.4 times larger than its predecessor by parameter count, and researchers trained it to target forecasts to specific weather data stations — useful both for granular predictions and for evaluating the model against real, ground-truth data.

The model can forecast more frequently in part because it ingests weather satellite data collected in real time on an hourly basis, rather than relying only on the processed datasets produced by government weather agencies. Google says WeatherNext 3 is the “first” AI model to directly incorporate raw observations for a high-resolution global forecast, though AI weather startup WindBorne says its own model, WeatherMesh 6, has been incorporating raw observations from its fleet of weather balloons and other sources since late 2025. Google responded that its own forecasts are higher resolution across the globe; both models still rely on national weather datasets to some degree.

Faster, cheaper AI forecasts also promise to matter most where accurate weather data has historically been out of reach: DeepMind staff research scientist manager Ferran Alet said higher-resolution forecasts of wind, rain, and cloud cover could make renewable energy projects more dependable, and better forecasts have already been cited as a way to improve crop yields in developing countries that lack expensive sensors and supercomputers.

Why it matters

The upgrade shows how fast AI is replacing physics-based supercomputing in weather forecasting, and because the gains come from cheaper compute rather than expensive sensor networks, they could matter most in the poorer, sensor-scarce regions where accurate forecasts have been hardest to get.