Google DeepMind Releases WeatherNext 2, a Forecasting AI Model for Users

Google DeepMind and Google Research introduced a new weather forecasting artificial intelligence (AI) model on Monday. Dubbed WeatherNext 2, it is the successor of last year's GenCast model and comes with several improvements. The Mountain View-based tech giant states that the model is now eight times faster and features an increased resolution (enabling detailed weather monitoring) for one hour. This is also the first time the company is bringing the AI model outside of its research labs and letting users access its capabilities.

DeepMind's WeatherNext 2 Released to Users

In a blog post, the tech giant detailed the new weather forecasting model and highlighted its availability. WeatherNext 2's data can now be accessed from Earth Engine and BigQuery. It is also being made available to Google Cloud's Vertex AI platform as part of an early access programme. Notably, the company already uses the technology in Search, Gemini, Pixel Weather, and Google Maps Platform's Weather application programming interface (API).

Put simply, WeatherNext 2 can generate hundreds of possible weather scenarios from a single input in less than a minute, using just one Tensor Processing Unit (TPU), Google's custom chip for AI work. This is a big deal because traditional weather models, which rely on physics-based simulations, often take hours to run on supercomputers.

The WeatherNext 2 shows measurable improvements, Google claims. It is said to outperform GenCast on 99.9 percent of key variables (like temperature, humidity, wind) and across lead times up to 15 days. It also delivers higher temporal resolution, meaning it can provide more detailed, hour-by-hour forecasts.

The technology behind the AI model is based on a new architecture called a Functional Generative Network (FGN). Rather than just producing a single line-of-best-guess forecast, FGN injects structured “noise” into the model's parameters so it can generate a variety of realistic and coherent weather futures.

The model learns what meteorologists call “marginals” and “joints”. Marginals are simple, separate variables, such as temperature or wind speed in a place. Joints are combinations, such as how wind and humidity interact over a whole region. Even though WeatherNext 2 is trained only on marginal data, DeepMind says it excels at forecasting joints too, which is important for complex weather patterns like heatwaves or storms.

Notably, WeatherNext 2 runs four times a day, producing six-hour forecasts each time.