WeatherNext 3 forecasts rain from satellites, not simulations

Google DeepMind released WeatherNext 3, and the change that matters most is where the data comes from. Older AI weather models, including WeatherNext 2, learn from the output of traditional physics simulations, which are recomputed only every six hours. WeatherNext 3 instead reads live global geostationary satellite observations directly, so it drops that six-hour lag and refreshes every hour. It also predicts surface conditions at 5-kilometer resolution, about five times sharper than the 25-kilometer grid of the previous model.

The accuracy numbers are concentrated in precipitation, which is where forecasts usually fail people. DeepMind reports a 60% improvement in CRPS, a standard probabilistic error score, measured against NASA's IMERG satellite rainfall data, plus a 30% gain against MRMS and roughly 10% against ground rain gauges at early lead times. Looking a day or more ahead, it claims up to 50% more accurate precipitation forecasts, with the biggest gains in regions that have few weather stations. The model also outputs wind speed at 100 meters, the height of a turbine, and high-resolution cloud and solar data. WeatherNext 3 is already feeding Google Search, Gemini, and Maps, with raw data available through BigQuery and Earth Engine.

Why it matters

If you plan around weather, a grid operator balancing wind and solar, a logistics team, or a farmer in an area with sparse stations, hourly satellite-driven forecasts change what you can rely on. The turbine-height wind and solar radiation outputs in particular are aimed squarely at renewable energy scheduling, so it is worth checking against your current forecast provider before the next planning cycle.