Google DeepMind Unveils WeatherNext 3 AI Weather Forecasting Model

Google DeepMind and Google Research have introduced WeatherNext 3, a new artificial intelligence system designed to deliver faster, higher-resolution weather forecasts across the globe. The model represents an upgrade over traditional forecasting approaches by using deep learning to predict atmospheric behaviour at finer detail and with greater frequency.

What Happened

Researchers at Google DeepMind and Google Research announced WeatherNext 3, an artificial intelligence weather model capable of producing forecasts down to a 5-kilometre resolution. In benchmark tests conducted on Operational WeatherBench—an evaluation tool created by the startup Brightband—WeatherNext 3 ranked as the most accurate model among leading competitors. It surpassed deep learning systems from Microsoft, Nvidia, and the European Centre for Medium-Range Weather Forecasts (ECMWF), as well as conventional supercomputer-driven systems operated by the ECMWF and the US National Weather Service.

Google confirmed that WeatherNext 3 will begin powering weather details displayed across Google Search, Google Maps, and Gemini, while also being made accessible to researchers and enterprise users through Google Cloud platforms. According to Google senior staff engineer Samier Merchant, this initiative marks the first time that several core atmospheric variables will directly support standard Google consumer products.

Key Highlights

• Tighter Resolution: WeatherNext 3 can predict core variables at a 5-kilometre resolution, improving upon previous artificial intelligence weather systems that typically operated over broader grids of 15 to 25 square kilometres.
• Hourly Outputs: The model generates forecasts on an hourly basis, moving away from standard six-hour forecasting cycles.
• Improved Precipitation Metrics: Evaluations on rain prediction show a 60 percent improvement compared to its predecessor, WeatherNext 2.
• Expanded Architecture: The new architecture features 2.4 times more parameters than WeatherNext 2 and is trained to target measurements from specific ground weather stations.
• Raw Satellite Ingestion: The model incorporates raw real-time satellite data on an hourly basis rather than relying solely on pre-processed supercomputer datasets.

Why This Matters

Standard meteorological predictions depend on government supercomputers calculating complex equations of atmospheric physics. While accurate, those traditional computations demand extensive time and expensive hardware. In contrast, deep learning systems learn patterns directly from vast datasets, allowing predictions to be generated much more quickly and at a fraction of the computing cost.

Lower computational requirements could benefit developing regions where expensive supercomputers and sensor networks are unavailable. The source material notes that public figures such as Bill Gates have pointed to artificial intelligence forecasting as a means to improve crop yields in developing countries. Furthermore, Ferran Alet, a staff research scientist manager at DeepMind, highlighted that detailed predictions of wind, cloud cover, and precipitation can make renewable energy operations more reliable.

What to Watch Next

While Google described WeatherNext 3 as the first artificial intelligence model to directly integrate raw observations for a high-resolution global forecast, atmospheric startup WindBorne noted that its WeatherMesh 6 model has incorporated raw observations from weather balloons and other inputs since late 2025. Google pointed out that WeatherNext 3 maintains a higher global resolution. However, because both systems still rely partly on government weather datasets, ongoing development will focus on achieving true direct data assimilation across the sector.

Frequently Asked Questions

What products will use WeatherNext 3?

Google plans to integrate the model’s data into Google Search, Google Maps, and Gemini, as well as making it available to developers and researchers on Google Cloud platforms.

How does WeatherNext 3 compare to WeatherNext 2?

WeatherNext 3 features 2.4 times more parameters, improves rain prediction evaluations by 60 percent, increases forecast frequency to an hourly schedule, and narrows resolution down to 5 kilometres.

Does WeatherNext 3 replace conventional government models?

While WeatherNext 3 outperformed traditional models from ECMWF and the US National Weather Service on the Operational WeatherBench utility, it still depends on underlying government datasets to carry out forecasts.

Source: TechCrunch report on research announcements from Google DeepMind, Google Research, and Brightband.