Google launches WeatherNext 3, its most advanced AI

What happened

Weather shapes an enormous range of everyday and high-stakes decisions, from whether to carry an umbrella to how supply chains, farms, and power grids respond to storms, heatwaves, and droughts.

The company reports meaningful accuracy gains as a result, with medium-range forecasts showing improvement of as much as 60per cent against one satellite-based benchmark, 30per cent against a radar-based one, and 10per cent against ground rain-gauge measurements at early lead times.

Beyond faster, sharper forecasts, WeatherNext 3 adds predictions tailored to renewable energy planning — including turbine-height wind speeds and detailed cloud cover and solar radiation estimates — that grid operators and renewable developers can use to match clean-energy output with demand.

Google says the model's approach is especially significant for parts of Latin America, Africa, and the Asia-Pacific region that have historically had limited access to high-resolution forecasting, since running traditional regional weather models at that level of detail typically requires supercomputing resources few local agencies can afford.

WeatherNext 3 began powering weather features in Google Search, the Gemini app, Google Maps, the Google Maps Platform Weather API, and Google Earth Engine starting the day of the announcement. Google says forecasts a day or more out will be up to 50per cent more accurate on precipitation specifically, with the largest gains in regions where forecasting has traditionally been weakest.

Why it matters

While AI has already sped up and sharpened forecasting in recent years, predicting fast-moving, hyper-local weather has remained a weak spot for existing models, largely because they lacked fine-grained detail and struggled to fold in real-time data from sources like satellites. WeatherNext 3's central innovation is what it learns from.

Earlier models, including its predecessor WeatherNext 2, were trained primarily on output from traditional numerical weather prediction systems — physics-based simulations that run on supercomputers but carry roughly a six-hour data lag, introducing bias for quickly shifting variables like rainfall or surface temperature.

The new model instead ingests a continuously updating mosaic of live geostationary satellite data, allowing it to generate a fresh forecast every hour rather than in six-hour blocks. It also trains directly on sparse weather station readings, letting it capture the kind of sharp local variation seen near coastlines, valleys, and mountain ranges that older models tend to smooth over.

The resolution gains are substantial: WeatherNext 3 can resolve surface variables like temperature and moisture down to a 5-kilometer grid, other surface variables at 10 kilometers, and atmospheric variables such as wind speed at 25 kilometers — roughly five times sharper overall than WeatherNext 2's uniform 25-kilometer, 6-hour-increment forecasts. Independent live evaluations from Brightband currently rank it as the top-performing global weather model available.

Precipitation has long been one of the hardest things for weather models to get right, since rain and snow are driven by small-scale, fast-moving cloud processes that physics-based simulations often blur or miss entirely. To tackle this, Google trained the model on two high-quality precipitation datasets: NASA's satellite-based IMERG product and its own radar-based global precipitation reanalysis.

What to watch

Developers and researchers can also access the underlying hourly, high-resolution data directly through BigQuery, Earth Engine, or bulk downloads from Google Cloud Storage. Google notes that the model doesn't replace official sources for severe weather warnings or public safety advisories, which it says should still come from local meteorological agencies and national weather services.