Brandon Fletcher
In 2026, accurately predicting hyper-local weather remains a challenge, affecting various aspects such as supply chains, sporting events, tourism, and energy production. Google’s latest AI advancement aims to improve this situation significantly.
On Thursday, Google DeepMind and Google Research launched WeatherNext 3, which they tout as their most sophisticated and precise global AI weather model to date.
According to the company’s press release, “Our model learns directly from real-time observations, enabling it to deliver timely, localized predictions for weather events that matter most to people.” By utilizing raw satellite data, the model generates high-resolution forecasts every hour, making reliable predictions available across Google products worldwide.
The incorporation of real-time, hourly satellite imagery represents a significant breakthrough for this model.
To provide context, conventional weather forecasting relies on numerical weather prediction methods. These models gather atmospheric data from satellites, weather stations, and weather balloons, processing it through complex physics simulations on supercomputers to project future weather conditions.
In contrast, AI weather models can produce forecasts much faster but still largely depend on data from traditional systems. Google points out that this can lead to a delay of approximately six hours, complicating the prediction of rapidly changing weather scenarios.
WeatherNext 3 adopts a novel approach by integrating real-time satellite observations directly into the AI framework, in addition to traditional weather analysis data. It is also trained using sparse data from weather stations, enabling it to deliver a new global forecast every hour with significantly more localized predictions.
For certain surface variables, such as temperature and moisture, WeatherNext 3 offers forecasts with a resolution of up to 5 kilometers. This allows the model to illustrate how weather conditions differ in areas that are just five kilometers apart. Wind speed forecasts are made with a resolution of 25 kilometers.
This high level of detail is crucial in regions where weather can vary greatly over short distances, such as coastal areas and mountainous terrains.
Google claims that WeatherNext 3 can achieve up to 50% greater accuracy in precipitation forecasts when users look a day or more ahead.
This level of improvement is particularly significant for regions in Latin America, Africa, and Asia-Pacific, where high-resolution forecasting has previously been hampered by the prohibitive costs of the supercomputers required for traditional models.
WeatherNext 3 has already secured the top position on Operational WeatherBench, a respected independent leaderboard maintained by AI weather forecasting startup Brightband, which evaluates the accuracy of various AI and traditional weather models.
The model also includes forecasts specifically tailored for renewable energy production. It predicts wind speeds at 100 meters above ground level—approximately the height of a wind turbine—along with cloud cover and solar radiation data. This information can assist wind and solar operators in estimating the electricity output of their facilities.
Starting Thursday, Google announced that WeatherNext 3 will enhance weather features across Google Search, Gemini, Google Maps, the Google Maps Platform Weather API, and Google Earth Engine.
Despite these advancements, Google continues to recommend that users consult their local meteorological agencies or national weather services for official forecasts, severe weather alerts, and public safety announcements.
Google did not provide an immediate response to a request for further comments.
Daniel Mercer
Ethan Carter
