Google's WeatherNext 3 Forecasts Rain Hourly at 5 Kilometers

Google DeepMind released WeatherNext 3 on Sept. 3, an AI weather model that refreshes its forecasts every hour at a 5-kilometer resolution, about five times sharper than its predecessor's 25-kilometer grid.

The model is now live in Google Search, the Gemini app, Google Maps and the Google Maps Platform Weather API, Google said in its announcement, with developer access through BigQuery, Google Cloud Storage and Google Earth Engine.

WeatherNext 3 has about 2.4 times as many parameters as WeatherNext 2 and trains directly against satellite and radar data instead of relying only on physics-based simulations, a change Google said lets it capture sharp, localized rain bands that earlier models blur. Predicted variables now include 100-meter wind speed, aimed at forecasting output from wind turbines.

What changed in the numbers

Metric	WeatherNext 3	Prior model
Surface resolution	5 km	25 km (WeatherNext 2)
Forecast refresh	Hourly	Every 6 hours (WeatherNext 2)
Global 2m-temperature error, August	Lowest on 26 of 30 days	— (Brightband)

Google said the model's medium-range precipitation forecasts are up to 60 percent more accurate than its predecessor when checked against NASA's IMERG satellite data, and up to 30 percent more accurate against the MRMS radar network. Longer-range precipitation forecasts improved by up to 50 percent, with the largest gains in Latin America, Africa and the Asia-Pacific region, where Google said fewer ground weather stations exist to feed traditional models.

Where an independent test agrees

Brightband, an atmospheric-science firm that runs the Operational WeatherBench leaderboard, found WeatherNext 3 posted the lowest global 2-meter-temperature error of any model it tracks on 26 of 30 days in August, ahead of models from Microsoft, Nvidia, the European Centre for Medium-Range Weather Forecasts and the U.S. National Weather Service, according to TechCrunch. Daniel Rothenberg, an atmospheric scientist at Brightband, said training the model against what a specific weather station measures moves the forecasting task closer to what people actually experience, TechCrunch reported.

What a competitor does differently

WindBorne Systems' competing model, WeatherMesh 6, has used raw atmospheric observations since late 2025, TechCrunch reported, a design WindBorne adopted before Google did. Google said WeatherNext 3's advantage is resolution, covering the globe at up to 5 kilometers, wider high-resolution coverage than WindBorne offers, by Google's own comparison. Neither company's claim has been checked by a third party under matched test conditions.

Samier Merchant, a Google senior staff engineer, said WeatherNext 3 marks the first time core Google products will run on this kind of weather model directly, according to TechCrunch. Ferran Alet, a DeepMind staff research scientist, told TechCrunch that machine learning suits weather prediction because it targets the approximate, noisy physics the field is actually trying to solve, rather than a cleaner idealized version of it.

The release follows Google shipping a cybersecurity-specialized Gemini model gated to defenders and an image-editing tool inside Workspace earlier in September, part of a broader push to fold DeepMind research directly into consumer products rather than ship it only as a standalone research release.

Google did not disclose WeatherNext 3's exact parameter count, only that it is about 2.4 times larger than WeatherNext 2's, a figure Google has also never published. Brightband said it will keep scoring the model against rivals on its public leaderboard as more data comes in through September.