Google Deepmind is moving its weather AI work into a faster and more widely available phase with WeatherNext 2, a new version of its forecasting model. The company says the upgrade improves performance across 99.9 percent of all meteorological variables and forecast ranges compared with the previous release.
What WeatherNext 2 changes
WeatherNext 2 is designed to improve core weather predictions across familiar measurements such as temperature, wind, and humidity. The model covers timeframes from zero to 15 days, which places it in the range people commonly rely on for daily planning, travel decisions, and short-term risk awareness.
The most direct claim from Google is speed. WeatherNext 2 can produce forecasts eight times faster, and it can generate outputs at resolutions as fine as one hour. That matters because weather is not a single fixed answer. Forecasting often depends on comparing many possible outcomes and understanding how conditions may evolve.
Google says the model can generate hundreds of possible weather scenarios in under a minute on a single TPU. By comparison, traditional physics-based systems running on supercomputers would need hours to complete the same task.
Why scenario speed matters
The useful part of faster weather AI is not only that a forecast arrives sooner. It is that many versions of a forecast can be produced quickly enough to compare. When a system can generate hundreds of possible scenarios, it can give forecasters and products more information about how weather might change within the same overall window.
Based on the source information, the main gains are concentrated in three areas:
- Coverage: WeatherNext 2 works across meteorological variables and forecast ranges.
- Speed: Google says forecasts are produced eight times faster than with the previous release.
- Resolution: Outputs can be generated with detail as fine as one hour.
Those points are important because weather data has to be useful at different scales. A broad outlook helps with longer planning, while finer time resolution can make short-term changes easier to understand.
The technique behind the model
Deepmind links the improved performance to a new method called a Functional Generative Network. The source describes this technique as injecting perturbation signals directly into the model architecture.
The goal of that design is to keep predictions physically realistic. In plain terms, the model is not only generating possible weather outcomes. It is also being guided so those outcomes stay within plausible physical behavior.
That point is central to weather AI. Forecasting is not simply a pattern-matching problem. Weather systems have relationships between variables such as temperature, wind, and humidity, and a useful model has to handle those relationships in a coherent way.
Where users may see WeatherNext
WeatherNext is already built into several Google products and services. The source lists Google Search, Gemini, Pixel Weather, and the Weather API as current integrations.
Google Maps integration is also on the way. That would extend the reach of the model into another service where weather can affect how people interpret places, routes, and conditions.
The product list matters because WeatherNext 2 is not being presented only as a research result. It is part of a broader push to bring AI weather forecasts into tools people already use.
Part of a longer Deepmind weather effort
WeatherNext 2 follows years of AI-driven weather research at Deepmind. The source also notes that in December 2024, the lab introduced GenCast, a diffusion-based model intended to improve short-term and medium-range forecasting.
Together, WeatherNext 2 and GenCast show the same general direction: Deepmind is applying generative AI methods to forecasting problems where speed, uncertainty, and physical realism all matter. The key claim for WeatherNext 2 is that it advances the earlier WeatherNext model while making forecasts faster and more detailed.
For Google, the practical test will be how these improvements appear inside everyday products. For users, the promise is clearer weather information across common services, supported by a model that can evaluate many possible scenarios much more quickly than traditional systems described in the source.