Google has introduced WeatherNext 2, an updated AI-driven weather forecasting model designed to deliver faster and more precise predictions across its services. While the company highlights internal performance gains, the broader significance lies in how this system reflects the ongoing shift from traditional physics-based forecasting toward machine-learning models that can scale more efficiently across global platforms. WeatherNext 2 replaces the first-generation model while keeping the earlier version accessible for research and benchmarking.
According to Google, WeatherNext 2 improves accuracy across 99.9 percent of measured variables, covering temperature, wind behavior, precipitation, humidity, and atmospheric pressure. These gains stem from a new modeling approach built around a Functional Generative Network, which replaces the Graph Neural Network and diffusion architecture used in WeatherNext 1. Instead of relying on step-based physical simulations, the FGN introduces controlled noise within the model to keep outputs realistic and internally consistent. This approach reflects an emerging trend in climate modeling, where machine-learning frameworks aim to reproduce complex atmospheric interactions without the heavy computational load of traditional numerical models.
Speed is one of the most notable changes. Google says the new system generates forecasts eight times faster than its predecessor, completing scenarios in under a minute when running on a TPU. In comparison, many physics-based models require hours on large computing clusters for similar output windows. WeatherNext 2 maintains the previous approach of issuing four six-hour forecasts each day and continues to support up to 15-day lead times, which keeps it aligned with the timelines expected in consumer weather apps.
These improvements feed directly into Google’s broader ecosystem. WeatherNext 2 is now supporting updated forecasts in Search, Gemini, the Pixel Weather app, and the Google Maps Platform’s Weather API. The expanded integration suggests Google is moving toward a unified forecasting pipeline that can serve both consumer-facing tools and developer platforms. The model is already available through Earth Engine and BigQuery, and it is being introduced to Vertex AI through an early-access program. WeatherNext 1 remains available for comparison, which may help researchers evaluate how machine-learning-based weather predictions evolve over time.
The shift to WeatherNext 2 reflects a growing industry interest in AI-generated weather data, not as a replacement for physics-based modeling, but as a complement that can deliver quick updates for real-time applications. Faster iteration can be valuable for mapping tools, travel planning, emergency alerts, and everyday forecasting. As WeatherNext 2 expands across Google Maps and other services, the model’s performance will likely be scrutinized for how reliably it captures rapid weather changes—an area where AI-based systems still face challenges compared to established meteorological methods.
