Google drops WeatherNext 2 with 8x faster weather predictions

Forecasting extreme weather has traditionally required supercomputers running heavy physics equations for hours on end.

Google DeepMind and Google Research just released WeatherNext 2, an AI model that cuts that compute time down to seconds while generating hundreds of possible forecast scenarios.

Running on a single TPU, the new model delivers predictions in under a minute, running eight times faster than its predecessor with temporal resolution down to a single hour.

Hundreds of scenarios in minutes

Meteorologists do not just need to know the most likely weather. They need to understand worst-case scenarios to prepare for natural disasters.

WeatherNext 2 tackles this by predicting hundreds of different weather trajectories from a single starting point. In benchmarks, it outperformed the previous WeatherNext across 99.9% of variables like temperature, wind, and humidity over 0 to 15-day lead times.

The secret behind this leap is an architecture called a Functional Generative Network (FGN). By injecting noise directly into the model in function space, it produces forecasts that stay physically coherent.

The training method is equally clever. Google trained the neural network solely on isolated weather elements known as marginals, like temperature at a single spot or wind speed at an altitude. From there, the system learned to forecast complex, interconnected systems, such as heat domes or entire wind farm energy yields.

Straight from the lab to Google Maps

Google is not leaving this research sitting in an academic paper.

Forecast data is already accessible inside Google Earth Engine and BigQuery. Enterprise developers wanting custom inference can also sign up for an early access program on Google Cloud Vertex AI.

Everyday users are already touching it too. The technology is now powering forecasts in Google Search, Gemini, Pixel Weather, and the Maps Weather API, with Google Maps app integration rolling out in the coming weeks.

Google previously used experimental versions of this model to help meteorological agencies map cyclone paths.

Why it matters

Weather prediction is quietly undergoing one of the fastest AI transformations of any scientific field.

For businesses in energy, logistics, and aviation, this is a massive operational upgrade. Instead of waiting hours for a supercomputer to spit out a single physics-based simulation, operators can now spin up hundreds of ensemble outcomes in minutes on standard cloud hardware to manage supply chain risks or wind farm outputs.

We think Google’s rollout strategy here is brilliantly calculated. Rather than simply open-sourcing a model checkpoint, DeepMind hooked the data pipeline straight into BigQuery, Earth Engine, and Vertex AI. It turns scientific prestige into immediate enterprise cloud lock-in.

The lingering question is whether data-driven generative models can hold up when unprecedented climate anomalies strike. If a warming planet triggers weather dynamics never seen in historical training sets, relying purely on neural nets over traditional atmospheric physics will face its real stress test.