Five Times Sharper, Every Hour: Inside WeatherNext 3, Google DeepMind's New Champion Weather Model

On September 3, Google DeepMind and Google Research quietly shipped what may be the most consequential applied-AI release of the week: WeatherNext 3, a new global weather forecasting model that produces a fresh forecast every hour at up to 5-kilometer resolution — roughly five times sharper than its predecessor — while cutting rainfall forecast errors by up to 60%. And unlike a typical research demo, it went straight into production: starting September 4, WeatherNext 3 powers the weather information inside Google Search, Maps, the Gemini app, the Google Maps Platform Weather API, and Google Earth Engine, with raw data available via BigQuery, Earth Engine, and bulk download from Google Cloud Storage.

Why this matters more than it sounds

Weather forecasts influence billions of decisions every day — some trivial (grabbing an umbrella), some anything but. Wind, rain, heatwaves, and droughts cascade through agriculture, supply chains, clean-energy production, and national economies. For decades, those forecasts came from government supercomputers laboriously solving physics equations — expensive, slow, and updated only every six hours on grids of 15–25 kilometers.

Since the European Centre for Medium-Range Weather Forecasts (ECMWF) opened up half a century of weather data in 2018, deep-learning models have been closing the accuracy gap at a fraction of the compute cost. WeatherNext 3 is the moment that progression stops being incremental: according to independent real-time evaluation by the startup Brightband on its Operational WeatherBench, it is now the most skillful medium-range global forecast model in the world — beating not only other AI models from Google, Microsoft, Nvidia, and ECMWF’s own AIFS, but also the traditional physics-based forecasts from the U.S. National Weather Service and ECMWF. During August, its 2-meter temperature forecasts had the lowest error of any peer on 26 of the last 30 days.

Three technical leaps

1. It learns from satellites, not supercomputers. Most AI weather models — including WeatherNext 2 — are trained on outputs of numerical weather prediction (NWP) systems, which carry a six-hour data lag that biases fast-changing variables like rain and surface temperature. WeatherNext 3 instead ingests a live mosaic of geostationary satellite observations every hour, making it what Google calls the first AI model to directly incorporate raw observations for a high-resolution global forecast. (The AI weather startup WindBorne notes its WeatherMesh 6 has been assimilating balloon observations since late 2025 — though at lower global resolution; both still lean on national datasets, so fully direct assimilation remains an open frontier.)

2. Station-level targets, not just grid averages. Traditional models output variables averaged across a 3D grid. WeatherNext 3 also predicts readings at specific weather stations — “what Denver’s airport’s weather station is going to measure on an hourly basis,” as Brightband atmospheric scientist Daniel Rothenberg put it. That grounds the model against real ground-truth data and resolves the dramatic temperature and humidity swings that occur over just a few kilometers near coastlines, valleys, and mountain ranges. The team also trained directly on sparse station observations, bringing high-fidelity forecasting to Latin America, Africa, and Asia-Pacific regions historically underserved by prohibitively expensive regional models.

3. Purpose-built precipitation and energy variables. Rain is the hardest problem in forecasting — cloud processes operate on tiny scales that physics simulations smear into blurry averages. WeatherNext 3 was trained on two premium data sources: NASA’s satellite-based IMERG retrievals and Google’s own satellite-radar precipitation reanalysis. The result: CRPS improvements of up to 60% against IMERG, 30% against MRMS, and 10% against rain gauges at early lead times. Users planning a day or more ahead will see up to 50% more accurate precipitation forecasts in Google products. The model also forecasts 100-meter wind speeds (turbine height), high-resolution cloud cover, and solar radiation — variables designed for grid operators and renewables developers to match clean-energy supply with demand.

Under the hood

WeatherNext 3 is a single Functional Generative Network (FGN) mesh transformer with 2.4× more parameters than WeatherNext 2. It’s initialized every hour, runs a 64-member ensemble for probabilistic forecasts across a 15-day horizon, and outputs a hierarchy of resolutions: temperature and moisture at 5 km, other surface variables at 10 km, atmospheric variables like wind at 25 km. The same architecture natively emits dense gridded fields, discrete cyclone tracks, and those sparse station-level predictions.

“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 Google senior staff engineer, told TechCrunch. “Weather is chaotic, and so small differences really start to perturb massively,” added DeepMind’s Ferran Alet. “Machine learning targets the problem we are really solving, which is approximate noisy physics from incomplete information and finite compute.”

The bigger picture

The transformer revolution in meteorology has been as consequential as the LLM boom, just quieter. Four of the top five models on Brightband’s live leaderboard are now AI-based. European and U.S. weather agencies already fold AI forecasts into operational products, and their speed and low cost promise real economic impact in developing regions where supercomputers and dense sensor networks were never affordable.

There are caveats. Google itself directs users to national meteorological agencies for severe-weather warnings, and both WeatherNext 3 and WindBorne’s model still depend on national observation infrastructure to initialize. But the direction is unmistakable: the forecast on your phone is now generated by a neural network watching satellites in real time — and it just got dramatically better at telling you whether you’ll need that umbrella.