Google Rolls Out New AI Weather Model That Delivers Global Forecasts 8 Times Faster With Stronger Accuracy Across 99.9% Of Variables

Googles New AI Forecast Model Promises Faster, Sharper Weather Predictions

Google DeepMind is pushing its AI weather research into everyday use, introducing WeatherNext 2 — a new forecasting system designed to deliver faster, sharper, and more flexible predictions than traditional tools.

Rather than relying on full-scale physics simulations that often take hours to compute, the model produces detailed global forecasts in under a minute, allowing agencies and consumers to react more quickly to changing conditions.

Senior director of research Peter Battaglia said the upgrade is ready for real-world deployment after years of experimental work.

He told reporters,

“We’re taking it out of the lab and really putting it into the hands of users in more ways than we have before… we have confidence that our forecasts are really quite effective and quite useful.”

How Weathernext 2 Changes Forecasting

At the core of the system is a modelling approach known as a Functional Generative Network, or FGN.

Instead of learning entire weather systems as one package, the model is trained on individual components such as temperature, wind, humidity, and pressure at specific points.

From there, it learns how these elements form the larger patterns behind storms, heatwaves, or wind shifts.

FGN also introduces targeted noise into each prediction, allowing the model to generate hundreds of plausible outcomes from a single starting point without repeating long computational cycles.

DeepMind says the method raises accuracy across 99.9 per cent of tested variables and increases resolution, enabling one-hour steps rather than the six-hour increments used previously.

According to Battaglia, the system “is about eight times faster than the previous probabilistic model that we released last year, and in terms of resolution, it is six times greater”.

He added that the model outperforms its predecessor “on 99.9% of the variables that we tested”.

How Much Faster And Why It Matters

A full WeatherNext 2 forecast runs on a single Google Tensor Processing Unit and takes less than a minute.

Traditional physics-based forecasts depend on recreating atmospheric dynamics — a process that can take hours even on a supercomputer.

The speed boost means predictions can be refreshed more often, improving the ability to track rapid-moving systems or emerging extreme weather.

DeepMind says the model delivers improved forecasts across temperature, wind, humidity, and pressure for nearly every point in its 15-day window.

In accuracy tests using the Continuous Ranked Probability Score, the new system showed average improvements of 8.7 per cent for average-pooled CRPS and 7.5 per cent for max-pooled CRPS when measured against GenCast, the company’s previous diffusion-based model.

Stronger Performance For Cyclones

One of the system’s notable upgrades is its ability to track tropical cyclones more accurately.

When compared with the International Best Track Archive for Climate Stewardship, WeatherNext 2’s ensemble-mean forecasts reduced position errors by the equivalent of roughly 24 hours of lead time between three- and five-day predictions.

Even a slower 12-hour-timestep version of the model remained more accurate than GenCast beyond two days.

Early cyclone-prediction tools powered by this architecture have already been shared with weather agencies for testing.

Product manager Akib Uddin highlighted the practical benefits:

“You get more accurate forecasts, and you get them faster, and that helps everyone make the right decisions, especially as we start seeing more and more extreme weather.”

Where People Will See The Upgrade

WeatherNext 2 is already active in Google Search, Gemini, Pixel Weather, and the Google Maps Weather API.

Wider rollout across Maps is planned, along with an early-access programme for companies wanting custom modelling.

Forecast data is also available through Google Earth Engine and BigQuery for geospatial and large-scale analysis.

Uddin said integration across Google platforms is underway:

“Whether you're on search, Android, or Google Maps, weather affects everyone, and so by making better weather predictions, we're able to help everyone.”

Why Businesses Are Paying Attention

The one-hour resolution is proving valuable for energy operators, agricultural planners, logistics networks, and other sectors where small timing differences can shift output or cost.

Uddin explained,

“It helps them make more precise decisions relating to things that affect their business.”

By generating many coherent scenarios instead of a single deterministic outcome, the model aims to help organisations plan for risk rather than react to surprises.

Instead of simply offering a probability, WeatherNext 2 shows a spread of physically realistic outcomes, providing clearer insight into uncertainties during volatile conditions.

Google joins a growing field of groups working on AI-driven forecasting, including ECMWF, Nvidia, and Huawei.

But with WeatherNext 2 now embedded across widely used products, Google’s model may become one of the most visible examples of how AI is reshaping everyday weather information.