Google DeepMind launches WeatherNext 3 with hourly, sharp forecasts

Deciding whether to pack an umbrella has always felt like a guessing game. Traditional weather apps might tell you it's raining in your city, but your street stays bone dry while the next neighborhood is flooded.

That disconnect happens because local weather changes fast, while supercomputer physics simulations lag by around six hours.

Google DeepMind and Google Research launched WeatherNext 3, which independent benchmarking firm Brightband ranks as the most advanced and accurate global weather model today.

Crucially, this is not a locked-up research paper. The model rolled out immediately across Google Search, Maps, the Gemini app, and Google Earth Engine.

Ditching the 6-hour lag for live satellite feeds

Most AI weather systems, including WeatherNext 2, learn from the outputs of numerical weather prediction models. While useful, those supercomputer runs take hours to finish, meaning the forecast is already operating on outdated atmospheric information.

WeatherNext 3 takes a fundamentally different path. It bypasses that simulation bottleneck by training directly on live, hourly geostationary satellite mosaics and ground weather stations.

Because it reads real-time observations as they happen, WeatherNext 3 generates a fresh global forecast every single hour.

Spatial resolution sees an enormous jump as well. WeatherNext 2 pushed predictions on a 25-kilometer grid every six hours. WeatherNext 3 maps key surface conditions like temperature and moisture at a native 5-kilometer resolution, making the picture roughly five times sharper.

Other surface variables resolve at 10 kilometers, while upper-atmosphere wind speeds sit at 25 kilometers. For coastlines, valleys, and mountain ranges where local conditions swing wildly over short distances, that fine grain prevents the over-smoothed, pixelated approximations of older models.

Dialing in rain and clean energy tracking

Rain has always been the nemesis of atmospheric modeling. Cloud dynamics happen fast at tiny scales, leaving standard AI forecasts looking like blurry smears that fail to pinpoint where a storm actually hits.

To fix this, Google fed WeatherNext 3 high-quality data from NASA's IMERG satellite system alongside proprietary satellite radar reanalysis.

The resulting leap is substantial. In medium-range forecasts, the Continuous Ranked Probability Score against IMERG improved by up to 60%, alongside a 30% jump on MRMS and 10% gains against rain gauge ground truth. For users checking plans a day or more in advance, precipitation accuracy improves by up to 50%.

DeepMind also added features specifically targeted at the green transition. The model outputs 100-meter wind speeds, which sits right at wind turbine hub height, as well as precise surface solar irradiance. Grid operators and solar farms can now forecast actual power yields far more reliably.

Why it matters

We think WeatherNext 3's real impact lies in democratizing high-resolution forecasting.

Historically, running localized 5-kilometer weather grids required massive regional supercomputing clusters that only wealthy meteorological agencies could afford. Regions across Latin America, Africa, and parts of Asia-Pacific were left relying on coarse global models. By serving 5-kilometer hourly data globally via BigQuery and Google Maps, DeepMind effectively levels that playing field.

For everyday users, weekend planning gets far more dependable, while developers get instant access to clean API feeds without wrestling with raw NetCDF files or heavy infrastructure.

Tying the model directly to live satellite feeds rather than intermediate physics simulations is an exceptionally smart architectural choice that eliminates cumulative simulation drift.

That said, trading physics engines for pure data-driven transformers comes with caveats: the model is only as reliable as satellite sensor calibration, and how gracefully it handles once-in-a-century climate anomalies remains an open question.