Google DeepMind and Google Research released a new AI weather model called WeatherNext 3 on September 3, 2026. The model is designed to predict weather patterns faster and with more detail than older tools.
Google plans to use the model inside several of its own products. This includes Google Search, Google Maps, and the Gemini app.
Samier Merchant, a senior staff engineer at Google, said this marks the first time some core weather variables will power many Google products at once. The company sees this as a major shift in how forecasts reach everyday users.
How the Model Performed in Testing
WeatherNext 3 was tested using Operational WeatherBench. This is a tool built by the startup Brightband that compares different AI weather forecasts.
The model beat other AI systems built by Microsoft, Nvidia, and the European Center for Medium-Range Weather Forecasting. It also outperformed traditional forecasts from the U.S. National Weather Service.
Most standard weather forecasts come from government supercomputers. These machines run complex physics equations, which makes them accurate but slow and costly.
Since 2018, researchers have trained AI models on decades of weather data. These models learn patterns instead of solving physics equations directly, which lets them run much faster.
Ferran Alet, a staff research scientist manager at DeepMind, said weather is chaotic by nature. He explained that machine learning is suited to working with noisy, incomplete data rather than perfect physics.
What Makes WeatherNext 3 Different
Older AI weather models had three main weaknesses. They covered wide areas instead of specific locations, struggled with rain predictions, and depended on formatted government data.
WeatherNext 3 aims to fix all three problems. It can now forecast down to a 5 km resolution for key weather variables.
Rain forecasts improved by 60% compared to WeatherNext 2, according to Google researchers. The model can also produce forecasts every hour instead of every six hours.
The new version has 2.4 times more parameters than its predecessor. Researchers also trained it to target specific weather stations, like an airport, for more precise local predictions.
Daniel Rothenberg, an atmospheric scientist at Brightband, said this approach connects the forecasting process closer to real-world use. He said predicting conditions at a specific station makes the tool more practical.
The model can update more often because it uses satellite data collected in real time. Google says this makes it the first AI model to use raw satellite observations for a high-resolution global forecast.
However, the startup WindBorne says its own model, WeatherMesh 6, has used raw observations from weather balloons since late 2025. Google responded by saying its own forecasts still cover the globe at a higher resolution.
Both models still rely on national weather datasets to complete their forecasts. Full independence from those government data sources has not been reached yet.
AI weather models are already used by weather agencies in Europe and the United States. Supporters say the lower cost of these tools could help regions that cannot afford expensive weather sensors or supercomputers.
Bill Gates has pointed to AI weather forecasting as a way to improve crop yields in developing countries. Alet added that better wind and rain forecasts could also help make renewable energy projects more reliable.
Google says the model is now available to researchers and businesses through its cloud platforms.
