DeepMind's New Storm Model Buys Forecasters One Extra Day, and Nobody Quite Knows Why It Works
WeatherNext Cyclones predicts where a tropical storm goes and how strong it gets, in one system, using data a hundred times coarser than specialist models. Code and weights are public.
Google DeepMind has published WeatherNext Cyclones, a forecasting model for tropical storms that sees roughly one day further ahead than the best operational systems in use today. To put that in perspective, one extra day of warning is about what conventional weather modelling gained over the previous decade. The work appeared in Nature and was built together with the US National Hurricane Center, the Cooperative Institute for Research in the Atmosphere, and the UK Met Office.
Cyclone forecasting has always involved an annoying trade. Global models such as the European ensemble system are good at predicting where a storm will go but too coarse to say how strong it will get. Specialist regional models like NOAA’s HAFS nail intensity but drift on the track. WeatherNext Cyclones does both in one system. Over a five-day forecast its estimated storm centre is off by an average of 230 kilometres, against 370 for the European ensemble and 335 for DeepMind’s own earlier GenCast model. On three-day intensity it beats HAFS by about 3.75 knots. During Hurricane Melissa, which hit Jamaica in 2025, DeepMind says the model helped forecasters catch the storm’s rapid intensification in time.
The genuinely odd part is the resolution. The model works on a grid where each point covers about 28 kilometres, roughly a hundred times coarser than the specialist regional models it outperforms. Even a compact variant at 111 kilometres per grid point stays competitive. “High resolution is not a strict prerequisite for state-of-the-art intensity forecasting,” the authors write. Their honest conclusion is that coarse atmospheric data apparently carries far more information about storm strength than anyone assumed, and exactly how the model extracts it is still an open research question.
Two technical choices did most of the lifting. Instead of the diffusion approach GenCast used, which needs dozens of passes through the network per step, this model uses Functional Generative Networks that get by with a single pass, making it eight times faster. And rather than adding random noise pixel by pixel, it injects noise into the network’s control layers, which produces genuinely different but physically plausible scenarios instead of fuzzy variants of one answer. Because a 15-day forecast now runs in under a minute on a Google AI chip, the team raised the number of parallel scenarios per storm from 50 to 1,000, which is what makes rare extreme outcomes show up at all.
What this means for you: Unless you live in a storm zone, this is not something you will touch. It matters anyway, for two reasons. First, it is a concrete case of AI producing a real public-safety gain rather than a benchmark win, and the code and model weights are on GitHub under an open licence, with a small version that runs free in a Colab notebook. Second, DeepMind is careful to frame it as a complement, not a replacement: the Hurricane Center still blends several physics-based models, and for official warnings you should still listen to your national weather service. That restraint is worth noticing in a field that usually oversells.
Sources
Source: https://deepmind.google/blog/weathernext-ai-model-achieves-breakthrough-in-forecasting-cyclones/
A Climate Scientist Logged Eight Weeks of AI Agent Use. It Drew 600 Times More Power Per Prompt Than a Chat Message
Zeke Hausfather tracked 3.2 billion tokens through Claude Code and estimated the electricity behind them. The per-prompt figures the AI labs publish turn out to describe a kind of use almost nobody does anymore.
