Author HubJonathan Kemper

Google Deepmind introduced WeatherNext 2, an upgraded version of its AI weather model that the company says outperforms the previous release across 99.9 percent of all meteorological variables and forecast ranges. The system delivers stronger results for core measurements like temperature, wind, and humidity for timeframes from zero to 15 days. According to Google, it also produces forecasts eight times faster and can generate outputs with resolutions as fine as one hour. The model can run hundreds of possible weather scenarios in under a minute on a single TPU, while traditional physics-based systems running on supercomputers would need hours to complete the same task.

Deepmind attributes the model's performance to a new technique called a Functional Generative Network, which injects perturbation signals directly into the architecture to keep predictions physically realistic. WeatherNext is already built into Google Search, Gemini, Pixel Weather, and the Weather API, and Google Maps integration is on the way.

Deepmind has been pushing hard on AI-driven weather research for years. In December 2024, the lab introduced GenCast, a diffusion-based model designed to further improve short-term and medium-range forecasting.