WeatherNext 3 sharpens Google's rain and snow forecasts
Google's new model resolves core variables at 5 km, refreshes them hourly, and reports a 60 percent gain over WeatherNext 2 on precipitation.
Symbolic image: a radar dome turns above a weather station as a technician adjusts a sensor arm in the drizzle.
Google has rolled out WeatherNext 3, an AI weather model that forecasts core variables hourly at 5 km resolution and, by the company's own numbers, predicts rain 60 percent better than its predecessor.
At a glance
- WeatherNext 3 was announced on September 3, 2026, built by Google DeepMind and Google Research.
- Core variables are resolved at 5 km and refreshed hourly, rather than on the usual six-hour cycle.
- The model carries 2.4 times as many parameters as WeatherNext 2.
- Scored on Operational WeatherBench, where Google places it ahead of Microsoft and Nvidia models and of NWS and ECMWF.
- Forecasts feed Google Search, Google Maps and Gemini, with data access through Google Cloud.
Google has shipped WeatherNext 3, the newest version of its AI weather model. It resolves core variables such as temperature, wind speed and humidity at 5 km and refreshes them every hour, rather than the six-hour cadence that is standard in operational forecasting. For rain, the company reports a 60 percent improvement over WeatherNext 2.
What changed under the hood
The model carries 2.4 times as many parameters as WeatherNext 2. Google DeepMind and Google Research built it jointly and announced it on September 3, 2026. Snowfall prediction is meant to improve alongside rain, and Google describes the output as reaching a resolution not achieved before.
The benchmark claim
Google evaluated the system on Operational WeatherBench. On that benchmark it says WeatherNext 3 beats AI models from Microsoft and Nvidia as well as its own earlier releases, and also comes out ahead of conventional forecasts from the U.S. National Weather Service and the European Centre for Medium-Range Weather Forecasts.
Every one of those figures comes from Google. No independent replication has been published so far, so how much of the advantage survives in day-to-day operations remains an open question.
Where the forecasts show up
Search, Maps and Gemini are the first consumer destinations, and Google Cloud carries the data for developers and researchers. Samier Merchant, a senior staff engineer at Google, framed the release as the first time several core variables feed directly into a wide set of Google products.
Physics versus learned patterns
Conventional forecasting integrates the equations of atmospheric physics on supercomputers, which is slow and expensive at fine resolution. Ferran Alet, a staff research scientist manager at DeepMind, argues that machine learning targets the actual task instead: approximating noisy physics from incomplete information.
For most people the change is invisible. The model swaps out in the background while the weather card in the app looks the same.
FAQ
What is Google WeatherNext 3?
It is Google's current AI weather model, from Google DeepMind and Google Research, announced on September 3, 2026 as the successor to WeatherNext 2.
How accurate is WeatherNext 3?
Google cites 5 km resolution on core variables, hourly updates and a 60 percent improvement on rain over WeatherNext 2. Those numbers have not been independently verified.
Where can I use WeatherNext 3 forecasts?
In Google Search, Google Maps and Gemini, with data access for developers and researchers through Google Cloud.
