Google's New AI Weather Model Updates Every Hour

Google DeepMind and Google Research released WeatherNext 3 on September 3, the company's most capable AI weather model yet, and the first designed to refresh its forecast every hour rather than every six. The model is already powering the weather results people see in Google Search, the Gemini app and Google Maps, with forecast data also flowing to developers through the Maps Platform Weather API, Google Earth Engine and Cloud Storage.

The headline change is resolution. WeatherNext 2 ran at a 25-kilometer grid and updated on a six-hour cycle, tied to the rhythm of government weather datasets. The new model predicts surface temperature at 5 kilometers, other surface variables at 10 kilometers, and atmospheric measurements such as wind at 25 kilometers. That is roughly five times sharper on the ground, and the hourly cadence means a forecast can react to a storm that forms mid-afternoon instead of waiting for the next scheduled run.

Reading the sky directly

What makes the hourly updates possible is a shift in what the model reads. WeatherNext 3 ingests live geostationary satellite mosaics as they arrive, rather than waiting for the processed datasets that national agencies publish a few times a day. Google describes the architecture as a Functional Generative Network built on a mesh transformer, and says the model carries about 2.4 times more parameters than its predecessor.

On the metric forecasters care about most, precipitation, Google reports up to 50% more accurate predictions a day or more ahead. Independent live evaluation by the forecasting firm Brightband rated it the most accurate global weather model currently running. As with any single-vendor benchmark, the numbers are worth treating as a starting point until more outside testing accrues, but the direction is clear enough.

Who actually uses this

The practical reach goes beyond a phone weather widget. WeatherNext 3 forecasts wind speed at 100 meters, about the height of a turbine, alongside cloud cover and solar radiation. Those are exactly the variables a grid operator needs to estimate how much power a wind or solar farm will produce in the next few hours, and getting them at hourly resolution is genuinely useful for planning around renewable supply.

The launch fits a wider pattern of AI models learning to anticipate physical events rather than just describe language. Earlier this year an MIT group showed a system that could forecast disasters it had never seen in the historical record. WeatherNext 3 is the more everyday version of that same idea: a model that has read enough of the atmosphere to guess, hour by hour, what it will do next.