Google DeepMind's WeatherNext 3 delivers hourly forecasts at 5km resolution

Google launches WeatherNext 3

Google DeepMind and Google Research have released WeatherNext 3, an AI weather model that produces hourly global forecasts at up to 5-kilometer resolution by learning from live satellite observations. The model moves on from the six-hourly, 25-kilometer output of its predecessor and is already feeding the weather information shown in Search, Maps, and Gemini, according to TechCrunch.

What has changed

As The Verge reports, WeatherNext 2 issued forecasts every six hours on a 25-kilometer grid. The new model generates a forecast each hour based on the latest satellite observations and can visualize variables such as temperature and moisture at up to 5-kilometer resolution, a global picture Google describes as five times sharper than before.

Under the hood, TechCrunch notes the model is larger than its predecessor, with 2.4 times more parameters, and its decoder heads were tuned to give more useful outputs. The designers also trained it to target specific weather stations rather than only averaging values across a three-dimensional grid, which allows both more granular predictions and direct evaluation against ground-truth measurements. The team has previously drawn praise for tuning its models to visualize cyclone paths.

The hourly cadence and finer resolution matter most for fast-moving systems that bring rain and snow. According to The Verge, using satellite data also fills gaps in regions with few ground-based rain gauges, largely outside the US and Europe, where Google expects the biggest forecast improvements.

The accuracy picture

On Operational WeatherBench, a comparison utility built by the startup Brightband that scores metrics such as temperature, wind speed, and humidity, WeatherNext 3 came out ahead of deep-learning models from Google, Microsoft, Nvidia, and the European Centre for Medium-Range Weather Forecasts, as well as traditional forecasts from the US National Weather Service and the ECMWF, TechCrunch reports.

The two sources give slightly different precision figures. The Verge says users will see precipitation forecasts up to 50 percent more accurate when looking at least a day ahead, while TechCrunch reports the model's rain evaluations are 60 percent improved over WeatherNext 2.

Google also designed the model to forecast renewable energy generation, including wind speed at 100 meters, roughly turbine height. DeepMind researcher Ferran Alet told The Verge that making renewable energy dependable matters as energy demand rises for Google and for humanity at large.

A disputed first

Google claims WeatherNext 3 is the first AI model to directly incorporate raw observations for a high-resolution global forecast. TechCrunch points out that the AI weather startup WindBorne says its WeatherMesh 6 model has been incorporating raw observations from its balloon fleet and other sources since late 2025. Google's response is that its forecasts are higher-resolution across the globe. Both models still rely on national weather datasets, so fully direct data assimilation remains future work.

Where it appears

WeatherNext 3 is now incorporated into Search, Maps, Gemini, and other Google products, and is available to users and researchers on Google's cloud platforms. The Verge reports Google has worked with the US National Hurricane Center and agencies in Asia to improve forecasting with its AI models.

The company is not declaring the end of physics-based forecasting. WeatherNext 3 is still trained on data from physics-based models, and weather agencies typically consult a range of predictions before issuing warnings, The Verge notes.

Why it matters

Traditional forecasting relies on supercomputers grinding through the equations of atmospheric physics, an approach that is accurate but expensive and comparatively slow. TechCrunch traces how the ECMWF's 2018 release of more than half a century of weather data opened the door for deep-learning models that predict quickly and at low cost, and notes that European and US agencies already fold AI models into their forecast products.

That speed and low cost could bring accurate forecasts to poorer regions where high-quality sensors and supercomputers have been out of reach, a point Bill Gates recently made when he cited AI-powered weather forecasting as a way to improve crop yields in developing countries. For Google, sharper forecasts of wind, rain, and cloud cover serve both the consumer products it ships daily and the renewable-energy planning its growing data-center footprint increasingly depends on.