WeatherNext 2 Launches With 8x Faster, High-Resolution Global Forecasting

Google DeepMind and Google Research have unveiled WeatherNext 2, described as the most advanced and efficient artificial intelligence model developed by the company for global weather forecasting. The new system has been positioned as a significant upgrade over the earlier WeatherNext model, with major improvements in accuracy, speed and scenario generation.

WeatherNext 2 has been engineered to deliver forecasts eight times faster than its predecessor, with each prediction being generated in under a minute on a single TPU, a task that traditionally takes hours on physics-based supercomputing models. Forecasts with hourly, high-resolution detail have also been enabled for the first time, expanding the model’s usefulness for sectors requiring precise short-term insights.

A new modelling method known as a Functional Generative Network (FGN) has been adopted, through which controlled noise is injected directly into the model architecture. This design has been credited with enabling the generation of hundreds of possible future weather scenarios from a single starting point, allowing meteorological agencies, planners and emergency management authorities to assess worst-case, best-case and most likely conditions more effectively. Despite being trained only on “marginals”—standalone weather elements such as temperature, humidity and wind—WeatherNext 2 has demonstrated improved ability to forecast “joints,” or complex, interconnected weather systems, which are regarded as essential for predicting heatwaves, rainfall patterns, and regional wind-energy outputs.

According to Google, WeatherNext 2 has surpassed the earlier model on 99.9% of atmospheric variables and across all 0–15-day lead times, marking a substantial enhancement in performance. Physically coherent and realistic predictions have been achieved at higher temporal and spatial resolution, supporting more reliable outcomes over longer ranges.

The technology has already begun transitioning from research to real-world operations. WeatherNext 2 forecast data has been integrated into Earth Engine, BigQuery, and made available through an early access programme on Google Cloud’s Vertex AI, enabling developers and enterprises to run custom model inference. Weather forecasts across Google Search, Gemini, Pixel Weather, and the Google Maps Platform Weather API have been upgraded using the new model, with Google Maps set to adopt the system in the coming weeks.

Google has stated that future plans include the integration of additional data sources, broader global access, and continued collaboration with researchers to support scientific discovery and climate-related decision-making.