Amid all the language-model news, it’s easy to forget that AI works elsewhere too. On August 10, Google DeepMind opened up its WeatherNext models – and they don’t predict the next chat token, they predict the weather.
What got released
Three models are now public: WeatherNext 2, a specialized WeatherNext Cyclones for tropical storms, and a smaller WeatherNext 2-mini. Google promises more accurate cyclone forecasts and ensembles that run up to 1,000 scenarios. There’s also a refreshed Weather Lab where you can view global weather and storm tracks in one place.
Ensembles are the key part for weather. Instead of a single forecast, the model runs many slightly varied passes and shows how confident – or shaky – a prediction is. A thousand scenarios give a finer picture of where a storm might actually land.
Why open matters here
Weather forecasting was long the domain of large government data centers running classic physics-based models. AI models like WeatherNext deliver comparable or better results – at a fraction of the compute. Google shipping the models in the open puts that into the hands of researchers, weather services, and developers who don’t run a supercomputer fleet.
With cyclones especially, every hour of lead time and every kilometer of accuracy counts. An open model that others can build on and adapt locally is more than a nice bonus there.
My take
It fits a pattern I keep seeing: the big labs aren’t releasing their crown jewels, but increasingly the specialized models that can move a whole field. On the same day, Meta shipped an open agent model – Google adds its weather AI to the pile.
For those of us who usually write about Claude and coding agents, it’s a healthy reminder: AI isn’t only chat. Sometimes it just buys lead time before the next storm.
Sources: