DeepMind's AI can forecast a cyclone a full day earlier — and it just open-sourced it
WeatherNext Cyclones, published in Nature this week, extends tropical-cyclone warning lead times by about 24 hours — roughly a decade of forecasting progress, on DeepMind's own numbers. The part that matters most isn't the accuracy. It's that the model is now public, so the national weather services that issue the warnings can actually use it.

Most AI announcements this year have been about models that write code or hold conversations. This one is about a model that could help decide when to evacuate a coastline.
Google DeepMind published WeatherNext Cyclones in Nature on 6 August 2026, alongside something more consequential than the paper: it open-sourced the model. The headline claim is that it adds more than a full day — about 24 hours — of lead time for predicting a tropical cyclone's track, intensity and wind structure. In DeepMind's framing, its three-day forecasts are now "as good as what prior models were able to provide for only the next two days," a jump it calls "roughly a decade's worth of meteorological progress." Those are DeepMind's own numbers, and they deserve independent replication before anyone treats them as settled — but the direction is not in doubt, and the reason it's credible is what the model has already done in the field.
Why a day matters more than it sounds
Cyclone forecasting has a specific, deadly failure mode: rapid intensification, when a storm jumps several categories in hours. It is the case traditional models handle worst and the one that kills people, because it collapses the window between "watch the weather" and "leave now." An extra 24 hours of reliable warning is not a marginal accuracy tweak; it is the difference between an orderly evacuation and a scramble.
DeepMind's approach is built for exactly that. Rather than produce a single forecast, WeatherNext uses what it calls Functional Generative Networks to generate a large ensemble of possible futures — scaled this year from 50 predictions to 1,000. A thousand plausible scenarios is how you catch the low-probability, high-consequence path where a storm suddenly explodes, instead of averaging it away. The model was trained end-to-end on nearly 20 terabytes of global atmospheric data and the IBTrACS database of nearly 5,000 historical storms.
It has already been used on a real storm
This is where WeatherNext separates from the pile of demo-grade AI. During the 2025 hurricane season, DeepMind says the model helped the U.S. National Hurricane Center make what it describes as a historic forecast for Hurricane Melissa — predicting the storm's rapid intensification and its landfall in Jamaica. That is a forecasting aid used by professional meteorologists on a live storm, not a lab benchmark. DeepMind is careful to say the obvious, and so should we: for official forecasts and warnings, the source remains your national weather service. The model informs the humans who issue the call; it does not replace them.
The open-sourcing is the real story
The accuracy gain is impressive. The distribution is what changes things. DeepMind has released WeatherNext 2 and WeatherNext Cyclones on GitHub, including a lightweight WeatherNext 2-mini that runs in a free Colab notebook. (For the broader argument about what "open" does and doesn't mean when a big lab ships model weights, we wrote a field guide.)
The significance is who this reaches. Frontier weather modelling has been the preserve of a handful of extremely well-resourced agencies with the supercomputers to run physics-based simulations. A capable cyclone model — one variant runs for free in a browser via Google Colab, no supercomputer required — puts serious forecasting within reach of the smaller, poorer, and often most storm-exposed national services that could never afford the traditional stack. If the field validation holds up elsewhere the way it did on Melissa, that is a rare thing in 2026: an AI release whose main effect is to hand a life-saving capability to the people who need it most, for free.
The honest caveats stand — the metrics are the vendor's, one hurricane season is not a track record, and independent testing across basins is the work that now matters. But strip those back and what's left is a real advance, already used on a real storm, now in everyone's hands. That is the version of AI progress worth being loud about.
Sources: Google DeepMind's WeatherNext announcement and the accompanying Nature paper (6 August 2026); the WeatherNext models on GitHub; DeepMind's write-up of the Hurricane Melissa forecast.
Ask Relay — he reads every question himself and replies personally by email.
