Hajra Naz
Scientists at Google DeepMind and Google Research released a major new AI model today. It’s designed for weather forecasting. This model sees changing atmospheric patterns more clearly. It also predicts atmospheric behavior more reliably than before.
The new model is called WeatherNext 3. It represents the latest wave in a broader transformation. Deep learning techniques have reshaped modern meteorology in recent years. Google says this model will soon power information users already see.
That includes weather data in Search. It includes Google Maps. It also includes Gemini. Beyond these consumer products, the model will be available more broadly too. Users and researchers can access it through Google’s cloud platforms.
“This is going to be the first time that some of the core variables feed and power a lot of the Google products,” said Samier Merchant, a senior staff engineer at Google.
This new model has already demonstrated strong performance. It proved to be the most accurate among leading contenders tested on Operational WeatherBench. This is a comparison tool built by the startup Brightband. It evaluates metrics like temperature, wind speed, and humidity.
WeatherNext 3 outperformed several other deep-learning models. That includes models built by Microsoft, Nvidia, and the European Center for Medium-Range Weather Forecasting (ECMWF). Beyond beating other AI systems, it also outperformed traditional forecasting methods. That includes forecasts from the U.S. National Weather Service. It also includes traditional ECMWF forecasts.
How AI Changed Weather Forecasting
Most weather forecasts historically come from government-owned supercomputers. These systems laboriously process complex mathematical equations. Those equations describe the underlying physics of weather patterns. While these systems have become remarkably accurate over time, they carry real drawbacks. They’re expensive to operate. They’re also comparatively slow.
A major shift happened back in 2018. The ECMWF released more than half a century of weather data. That data had been produced by these traditional physics-based systems. Deep learning researchers seized this opportunity. They began training models capable of making fast predictions. Notably, these predictions achieved accuracy comparable to traditional government tools.
“Weather is chaotic, and so small differences really start to perturb massively… Machine learning targets the problem we are really solving, which is approximate noisy physics from incomplete information and finite compute, and so it learns patterns from a lot of data,” said Ferran Alet, a staff research scientist manager at DeepMind.
Since that shift began, model developers have worked to address key weaknesses in AI forecasting. These systems often forecast over areas too broad to be truly useful, typically spanning 15 to 25 square kilometers. They haven’t always excelled at predicting rain accurately. They also still depend heavily on formatted datasets produced by government agencies.
How WeatherNext 3 Addresses These Challenges
WeatherNext 3 takes direct aim at all three of these challenges. On key variables, researchers explained, the model can predict down to a resolution of just 5 kilometers. Its rain predictions have improved significantly too. Evaluations show a 60% improvement over the previous version, WeatherNext 2. The model can also produce hourly forecasts now. That’s a major upgrade from the standard six-hour prediction interval.
These improvements stem from deliberate design choices. WeatherNext 3 is a considerably larger model. It contains 2.4 times more parameters than its predecessor. Researchers also tailored specific targets for the model’s decoder heads.
This helps generate more useful outputs overall. Most weather forecasts typically output metrics averaged across a 3D grid. DeepMind researchers have already earned praise for a specific enhancement, though. They tuned their model to visualize cyclone paths directly.
This latest version includes another notable innovation. Designers trained the model to target forecasts at specific weather data stations. This matters for two important reasons. It enables more granular, localized predictions. It also allows researchers to evaluate the model’s accuracy against specific, real-world ground-truth data.
“The idea, with a lot of AI applications, is to try to run tasks as end-to-end as possible,” said Daniel Rothenberg, an atmospheric scientist at Brightband. “Adding a capability where this model is now also predicting, say, what Denver’s airport’s weather station is going to measure on an hourly basis, just connects that forecasting task closer to the core.”
Real-Time Data Powers More Frequent Forecasts
This model can forecast more frequently for a specific reason. It can ingest weather satellite data collected in real time, updated hourly. Feeding AI models with raw empirical observations offers real promise.
This differs from relying solely on analysis produced by traditional weather supercomputers. This approach could lead to more accurate forecasts overall. Still, working with unformatted, raw data remains technically challenging.
Google describes WeatherNext 3 as the “first” AI model to directly incorporate raw observations. This applies specifically to high-resolution global forecasting. However, AI weather startup WindBorne offers a different perspective.
The company says its own model, WeatherMesh 6, has incorporated raw observations since late 2025. That data comes from WindBorne’s fleet of weather balloons and other sources. When asked about this comparison, Google emphasized something specific.
The company said its forecasts offer higher resolution across the entire globe. Regardless of this distinction, both models still rely on national weather datasets to generate forecasts. More work remains necessary to achieve true, direct data assimilation.
Broader Impact Beyond Consumer Weather Apps
While large language models attract most public attention, the transformer revolution in meteorology has proven equally significant. European and U.S. weather agencies are already incorporating AI models into their forecast products. These models offer real speed and cost advantages.
That combination promises meaningful economic impact in poorer regions. Historically, expensive sensors and supercomputers have kept accurate forecasts out of reach for these areas.
Bill Gates recently highlighted AI-powered weather forecasting as a crucial technological benefit. He noted that better forecasts can improve crop yields in developing countries specifically. Alet, the DeepMind researcher, pointed to another important application.
Higher-resolution forecasts covering wind, rain, and cloud cover could make renewable energy projects considerably more dependable.
“At the end of the day, I think Google is about providing useful information to the user, and a lot of what users are looking for has to do with the weather in some way or another,” Alet said.
