Google’s latest AI weather model gives you no excuse to forget your umbrella

Tim Fernholz reports that Google DeepMind and Google Research released WeatherNext 3, which will feed Search, Maps, and Gemini and sit on Google Cloud. Senior staff engineer Samier Merchant told TechCrunch this is the first time some of the core variables will power a lot of Google products. On Operational WeatherBench, a comparison utility from startup Brightband looking at temperature, wind speed, and humidity, the model is described as the most accurate among leading contenders, beating other deep-learning models from Google, Microsoft, Nvidia, and ECMWF, plus traditional forecasts from the U.S. National Weather Service and ECMWF.

Fernholz's technical extras: WeatherNext 3 has 2.4 times more parameters than its predecessor; it can predict key variables down to 5 kilometers; rain evaluations are 60 percent improved over WeatherNext 2; forecasts go hourly instead of every six hours. The model is also trained to target specific weather stations, not only grid averages. Brightband atmospheric scientist Daniel Rothenberg called station-level targeting a more end-to-end forecasting task. Google says it is the first AI model to directly incorporate raw observations for a high-resolution global forecast. WindBorne says its WeatherMesh 6 has ingested raw balloon and other observations since late 2025. Google's reply: its forecasts are higher resolution across the globe. Both still rely on national weather datasets.

This is the competitive version of the DeepMind blog. Brightband's Operational WeatherBench is the scoreboard Fernholz is willing to name, including wins over ECMWF and the NWS. That is a bigger claim than 'five times sharper than our last model,' and it should be checked on Brightband's board rather than in a sourced paragraph.

The WindBorne pushback is the paragraph Google did not write. 'First' is doing marketing work. Direct observation in is the research direction; it is not a trophy Google uniquely holds. Fernholz is correct that both systems still lean on national datasets, so 'raw satellite in, physics out' is incomplete.

Signal Desk's read: the 2.4x parameter bump and station heads are how Google bought resolution and a number it can eval against Denver's airport rather than a 25-kilometer cell. Putting those fields into Search is the actual launch. Beating ECMWF on a startup's live board is the press strategy. If WeatherNext 3 is as good as advertised, meteorological agencies will assimilate it or look late. If the board is a slice of variables, the umbrella headline is a consumer gloss on a grid-skill story.

Alet's line that weather is the information users already want is the corporate logic. Google does not need to become a weather service. It needs the weather module to stop being the place forecasts go to be wrong.

After ECMWF released decades of weather data in 2018, learned models started matching slow physics codes at a fraction of the cost. The remaining gaps Fernholz lists — coarse grids, weak rain, dependence on agency analyses — are exactly the three WeatherNext 3 is aimed at.