Google Research and Google DeepMind have released WeatherNext 3, the newest version of the company's AI weather forecasting system and, by Google's account, its most accurate global forecast model yet. Announced on September 3, the system will begin feeding weather information into Search, Google Maps and Gemini, and is available to businesses and researchers through Google's cloud platforms.

Unlike the government supercomputers that have produced most of the world's forecasts for decades, WeatherNext 3 is a deep learning model: it learns patterns directly from observations rather than step-by-step solving the equations of atmospheric physics. Google also says it is the first AI model to incorporate raw satellite observations directly into a high-resolution global forecast — a claim at least one competitor disputes. Here is what changed, and why forecasters well beyond Google are watching the latest AI developments in this field closely.

How AI Broke the Weather Forecasting Mold

Traditional forecasts come from national agencies such as the US National Weather Service and the European Centre for Medium-Range Weather Forecasts (ECMWF), whose supercomputers grind through mathematical equations describing the physics of the atmosphere. Those systems are remarkably accurate but expensive and comparatively slow.

The deep learning alternative took off after the ECMWF released more than half a century of archived weather data in 2018, giving researchers exactly the kind of training corpus neural networks thrive on. "Weather is chaotic, and so small differences really start to perturb massively," Ferran Alet, a staff research scientist manager at DeepMind, told TechCrunch. "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."

Sharper, Faster and Far More Frequent

WeatherNext 3 is a larger model than its predecessor — 2.4 times more parameters — and it now tops Brightband's Operational WeatherBench, an independent benchmark that compares AI forecast systems on temperature, windspeed and humidity. According to TechCrunch, it outperforms rival deep learning models from Google, Microsoft, Nvidia and the ECMWF, and beats the traditional numerical forecasts produced by the US National Weather Service and the ECMWF themselves.

The practical gains are substantial:

  • Resolution down to 5 km for key variables, where most AI models forecast across areas of 15 to 25 km. The Decoder reports the model is five times more detailed than WeatherNext 2.
  • Rain forecasting improved by 60 percent over WeatherNext 2 on Google's evaluations, and 9to5Google cites 50 percent more accurate precipitation forecasts for Google's own products.
  • Hourly forecasts instead of the standard six-hour cadence, because the model can ingest satellite observations collected in real time each hour.
  • Station-level targeting, meaning the model predicts what specific weather stations — not just grid squares — will measure.

"The idea, with a lot of AI applications, is to try to run tasks as end-to-end as possible," Daniel Rothenberg, an atmospheric scientist at Brightband, told TechCrunch. "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."

The Raw Satellite Data Dispute

Google's claim to a first rests on data plumbing as much as model architecture. Most AI forecasters still train and initialize on formatted analysis datasets produced by government weather agencies. WeatherNext 3 ingests raw satellite observations directly, hourly — closer, Google argues, to true data assimilation.

The AI weather startup WindBorne counters that its own WeatherMesh 6 model has been incorporating raw observations from its weather balloon fleet and other sources since late 2025. Asked about the dispute, Google pointed to the global resolution of its forecasts. Both companies, TechCrunch notes, still lean on national datasets for parts of the pipeline, so fully direct assimilation remains a work in progress for the whole field.

Where Users Will See It

"This is going to be the first time that some of the core variables feed and power a lot of the Google products," Samier Merchant, a senior staff engineer at Google, told TechCrunch. Weather information in Search, Google Maps and Gemini will draw on WeatherNext 3 output, putting AI-generated forecasts in front of billions of users — many of whom will never know the model behind them.

Why It Matters Beyond the App

The economics are the bigger story. AI forecasts are fast and cheap to produce compared with supercomputer ensembles, which puts usable forecasting within reach of regions that have never been able to afford it. The Decoder reports that parts of Africa, Latin America and the Asia-Pacific that lack accurate forecasts should see the largest gains. Bill Gates has cited AI weather forecasting as a concrete development benefit, with better forecasts improving crop yields in developing countries, and Alet notes that higher-resolution wind, rain and cloud forecasts can make renewable energy projects more dependable.

Meteorological agencies in Europe and the US are already folding AI models into their operational products. With WeatherNext 3, the transformer revolution that reshaped language has quietly become the engine of everyday weather.

Stay Ahead of AI

AI research moves fast. Get the latest AI developments in one place.

Read more AI news →