Google DeepMind Launches WeatherNext 3: Hourly AI Weather Forecasts at 5-Kilometer Resolution

Posted on 6th Sep 2026 08:27:13 in Artificial Intelligence, Machine Learning

Tagged as: AI, Google DeepMind, WeatherNext, weather forecasting, machine learning, climate tech

On September 3, 2026, Google DeepMind and Google Research launched WeatherNext 3, the most advanced global weather model the company has ever released — and the first one that generates forecasts every hour of the day. The model produces predictions at up to 5-kilometer resolution, roughly five times sharper than its predecessor's 25-kilometer grid, improves rain prediction by up to 60 percent a day out, and is already feeding the weather information shown in Google Search, Maps, and Gemini. On the Operational WeatherBench leaderboard, a neutral comparison utility built by the startup Brightband, WeatherNext 3 tops every leading contender — not only rival AI models from Microsoft, Nvidia, and the European Centre for Medium-Range Weather Forecasts (ECMWF), but the traditional physics-based forecasts of the U.S. National Weather Service and the ECMWF itself.

"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. For a field long dominated by government-owned supercomputers churning through equations of atmospheric physics, the launch marks the latest wave of a sea change that began when the ECMWF released more than half a century of weather data in 2018 and deep learning researchers started training on it.

From Every Six Hours to Every Hour

Traditional numerical weather prediction is a triumph of physics and engineering — and a slow, expensive one. Supercomputers solve equations describing atmospheric behavior and output forecast data in six-hourly cycles, which downstream systems then reformat for consumers. AI models have largely inherited that rhythm, along with its limitations: coarse grids of 15 to 25 square kilometers, notoriously weak rain predictions, and a dependency on the formatted datasets produced by government agencies.

WeatherNext 3 attacks all three weaknesses at once. It is an ensemble model that draws directly from raw satellite imagery collected in real time, allowing it to skip the standard six-hour lag between observation and forecast. That is the change that unlocks hourly forecasts — useful for tracking fast-moving conditions like rain and snow rather than waiting for the next six-hour cycle. The model carries 2.4 times more parameters than WeatherNext 2 and uses decoder heads tailored to different variables: temperature and humidity can be predicted down to 5 kilometers, targeted at specific weather stations, while other surface variables like wind run at 10 kilometers. Google says it is the first AI model to directly incorporate raw observations for a high-resolution global forecast.

Station targeting matters more than it sounds. By training the model to predict what a specific sensor — say, Denver International Airport's weather station — will measure hour by hour, researchers can evaluate against hard ground truth instead of grid-averaged estimates. "Adding a capability where this model is now also predicting what Denver's airport's weather station is going to measure on an hourly basis just connects that forecasting task closer to the core," said Daniel Rothenberg, an atmospheric scientist at Brightband. The result is the 60 percent improvement in one-day rain accuracy, the metric ordinary users feel most directly.

Outscoring the Supercomputers

Deep learning's advantage in weather is not magic but framing. "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. An AI model trained on decades of global observations can learn what the atmosphere actually does, without needing to model every physical equation behind it — and it can run in minutes on specialized hardware rather than hours on national supercomputers.

The track record behind WeatherNext 3 explains why forecasters are paying attention. In an August 2026 paper published in Nature, the WeatherNext team showed their cyclone model achieved state-of-the-art accuracy on storm track, intensity, and wind structure — with three-day forecasts as good as what prior models delivered for only two days. That extra day of lead time corresponds, the team calculates, to roughly a decade's worth of meteorological progress. During the 2025 hurricane season, WeatherNext helped the U.S. National Hurricane Center issue a historic advance warning for Hurricane Melissa, predicting the storm's rapid intensification and landfall in Jamaica five days ahead. Tropical cyclones — hurricanes and typhoons — have caused more than 700,000 deaths and $1.4 trillion in economic losses over the past 50 years, which is the human stakes behind every hour of warning gained.

Into Every Google Product — and Every Industry

WeatherNext 3's designers built it for deployment, not just for papers. The forecasts now flow into the weather cards billions of people see in Search, Maps, and Gemini, and the model is available to enterprises through BigQuery, Earth Engine, the Google Maps Platform, and Google Cloud Storage. For industry, Google is exposing variables that grids care about: wind, cloud cover, and solar radiation forecasts that help renewable energy operators plan output around the weather rather than against it. An interactive platform called Weather Lab, part of Google Earth AI, lets researchers explore live global forecast layers and track tropical cyclones in near real time against traditional meteorological baselines.

The economics could matter as much as the physics. Accurate forecasts have historically required the combination of dense sensor networks and supercomputing capacity that poorer regions could not afford; AI models that run cheaply and quickly could change that equation. To be clear about the current limits: WeatherNext 3 still relies on national weather datasets in addition to its raw satellite stream, and true end-to-end data assimilation remains a work in progress. The startup WindBorne has pointed out that its WeatherMesh 6 model has incorporated raw observations since late 2025, and Google concedes its own claim rests on being the first at this resolution and global scale. None of that diminishes the direction of travel: forecasting is becoming faster, sharper, and dramatically cheaper, and the race is now between AI labs and meteorological agencies working together rather than against each other — DeepMind's cyclone work was co-developed with the National Hurricane Center, CIRA, and the UK Met Office.

Sources

• Google DeepMind — WeatherNext 3: Our Most Advanced Global Weather AI Model
• TechCrunch — Google's Latest AI Weather Model Gives You No Excuse to Forget Your Umbrella
• AI Weekly — Google DeepMind Ships WeatherNext 3: Hourly 5km Global Forecasts, 60% Better Rain Accuracy
• Google DeepMind — WeatherNext: AI Model Achieves Breakthrough in Forecasting Cyclones