DeepMind's WeatherNext 3 Delivers Hourly AI Forecasts at 5-Kilometer Resolution

Weather forecasting has long been bottlenecked by the cadence of government-run models that refresh every six hours. On September 3, 2026, Google DeepMind announced WeatherNext 3, an AI-powered global weather model that changes that equation entirely — generating a new forecast every hour by drawing on live geostationary satellite observations.

What's New in WeatherNext 3

The jump from WeatherNext 2 to WeatherNext 3 is not incremental. The new model processes surface variables — temperature, moisture, wind — on a 5-kilometer grid, compared to the 25-kilometer resolution of its predecessor. That's a 25× finer spatial detail, delivered in hourly snapshots instead of six-hour blocks.

Accuracy improvements are just as striking. According to DeepMind's own benchmarks, rain and snow forecasts are up to 50% more accurate than previous models for lead times of one day or longer. The model also forecasts wind speeds at 100 meters above ground — roughly the hub height of a large wind turbine — along with cloud cover and solar radiation, data that grid operators can use to predict renewable power output more precisely.

How It Works

Traditional Numerical Weather Prediction (NWP) models rely on physics equations solved across global grids, fed by radiosonde balloon launches and satellite snapshots released every six hours by government agencies. WeatherNext 3 breaks that dependency by ingesting live geostationary satellite observations directly — data streams that are continuously available rather than released in fixed batches. The neural network was also trained on historical weather station observations, grounding its predictions in real-world surface measurements.

The result is a model that can update faster than any traditional NWP system, reacting to rapidly evolving conditions like flash floods or sudden wind shifts that a six-hour update cycle would miss entirely.

Where You'll Encounter It

Google says WeatherNext 3 will begin feeding into weather results across Google Search, Google Maps, and the Gemini assistant. It will also be available to researchers and enterprise customers through Google Cloud. Practically speaking, the next time you search "will it rain this afternoon," the hourly forecast you see may be coming from a neural network rather than a traditional weather model.

For energy companies and urban planners, the high-resolution wind and solar forecasts represent a direct operational tool — not just better predictions, but better grid dispatch decisions. A wind farm operator can now get an hourly update on expected wind speeds at turbine height rather than waiting six hours for the next model run.

The Bigger Picture

WeatherNext 3 is the latest in a line of AI weather breakthroughs that have accelerated sharply since 2023. DeepMind's earlier GraphCast model demonstrated that AI could outperform ECMWF's deterministic forecast on many metrics, and WeatherNext 3 now pushes that advantage into finer spatial and temporal granularity. The gap between AI-driven and traditional weather forecasting is widening faster than most meteorologists expected even two years ago.

Whether this eventually replaces government NWP infrastructure entirely, or serves as a high-resolution enhancement layer, is still an open question. For now, WeatherNext 3 gives everyone from weekend hikers to renewable energy operators a meaningfully sharper picture of what the atmosphere is about to do — updated every single hour.