Google launches WeatherNext 3 with 5 km resolution and hourly updates

Google DeepMind and Google Research have released WeatherNext 3, a weather model that outperforms every rival — including systems from Microsoft, Nvidia and the European Center for Medium-Range Weather Forecasts (ECMWF) — as well as the traditional forecasts of the US National Weather Service. For the first time, the core variables of an AI model will feed directly into Google Search, Maps, Gemini and the company's cloud platforms.

Behind the numbers

Researchers who spoke with TechCrunch say WeatherNext 3 reaches 5 kilometre resolution on key variables, a dramatic improvement over the 15 to 25 kilometres that characterised the previous generation of AI models. Precipitation accuracy improved 60% over WeatherNext 2, and forecast frequency jumped from once every six hours to a continuous hourly cadence. The model is 2.4 times larger than its predecessor by parameter count, and its decoder heads are tuned to produce actionable answers rather than averages across a three-dimensional grid.

How it works in practice

The most significant step is training the model to predict readings from specific weather stations, not just general spatial fields. Daniel Rothenberg, an atmospheric scientist at the startup Brightband who built the Operational WeatherBench comparison tool, explains that this approach brings the forecasting task closer to ground truth: when the model predicts what the Denver airport station will measure each hour, the forecast can be compared against a measured reality and improved accordingly.

The bigger picture

Traditional forecasts run on government supercomputers that solve physics equations — an expensive and relatively slow process. In 2018 ECMWF released more than fifty years of historical data, enabling deep-learning researchers to train models that learn patterns from vast datasets instead of solving equations. Ferran Alet, a research director at DeepMind, notes that weather is a chaotic system in which tiny differences amplify rapidly; machine learning attacks the real problem — noisy physics from partial data and finite compute.

What comes next

Samier Merchant, a senior engineer at Google, says this is the first integration of core variables from an AI model directly into the company's core products. The model can also ingest raw satellite data in real time at an hourly cadence, a capability that promises higher accuracy, though it still depends on the preprocessing performed by government agencies.