On 6 August 2026 Google DeepMind and Google Research published a paper in the journal Nature describing WeatherNext Cyclones, an AI weather model that forecasts where a tropical cyclone will go, how strong it will be and how wide its winds will reach. The headline result is time: the researchers report about one extra day of useful warning compared with the leading operational forecasting systems. On the same day Google released the model's code and weights so that researchers and weather agencies can run it themselves.
What happened
The research was announced in a post on the Google DeepMind blog, WeatherNext: AI model achieves breakthrough in forecasting cyclones, alongside the open-access paper Operational tropical cyclone forecasting with AI in Nature.
According to the paper's abstract, WeatherNext Cyclones produces ensemble forecasts (many alternative scenarios rather than a single prediction) of a cyclone's track, intensity and size, with scenarios running 15 days ahead. It was evaluated on cyclones from 2023 to 2025, and its track, intensity and wind-radius predictions gave an average lead-time advantage of one day or more over leading operational models.
DeepMind puts the same result another way: the model's three-day forecasts are about as accurate as earlier systems' two-day forecasts, which the team describes as roughly a decade's worth of meteorological progress. The work was done with forecasters at the US National Hurricane Center, the Cooperative Institute for Research in the Atmosphere at Colorado State University and the UK Met Office.
The technical change that matters most is that one model now does the whole job. Forecasters have traditionally relied on coarse global models for a storm's track and separate high-resolution regional models for its intensity. DeepMind says WeatherNext Cyclones predicts track, intensity and wind structure together, which removes that trade-off.
Key details
| Item | What the sources say |
|---|---|
| Published | 6 August 2026, in Nature (open access) and on the Google DeepMind blog |
| What it forecasts | Track, intensity and size of tropical cyclones, up to 15 days ahead |
| Main result | An average lead-time advantage of one day or more over leading operational models, tested on cyclones from 2023 to 2025 |
| Ensemble size | Up to 1,000 scenarios per forecast, compared with conventional 50-member ensembles |
| Training data | Nearly 20 terabytes of global atmospheric data plus a database of nearly 5,000 historical storms |
| Speed | A single 15-day forecast in less than a minute on a TPU chip, according to DeepMind |
| Release | Code and model weights for WeatherNext 2 and WeatherNext Cyclones, plus a smaller WeatherNext 2-mini that runs in a free Colab notebook |
The WeatherNext repository on GitHub lists the code under the Apache 2.0 licence and other materials under Creative Commons Attribution 4.0. It also carries some plain warnings: the models are experimental research, are not an officially supported Google product and are not endorsed by any government meteorological agency. The README says the full-size models need high-end accelerator hardware, while the lightweight Mini versions are meant for tighter memory and compute limits.
Why it matters
Cyclones are among the most expensive natural hazards. DeepMind's post cites more than 700,000 deaths and US$1.4 trillion in economic losses worldwide over the past 50 years. An extra day of reliable warning is time to evacuate, secure sites, move stock and reroute freight.
Three things stand out for anyone who follows AI rather than meteorology.
- It is already in operational use. DeepMind says WeatherNext helped the National Hurricane Center forecast the rapid intensification of Hurricane Melissa in 2025 and its landfall in Jamaica. This is AI that working forecasters have tested during real storms, not a laboratory benchmark.
- It is fast to run. A forecast that takes under a minute makes it practical to generate 1,000 scenarios, and the paper says large ensembles capture rare events better than the usual 50. Rare events are the ones that do the damage.
- It is open. DeepMind says the code and weights are freely available for anyone to build on, whether for academic research, operational forecasting or more specialised, localised models. ETV Bharat's report on the release quotes Google saying the aim is for researchers, non-profits and local weather agencies everywhere to build tools that better protect communities.
There is a limit worth stating. DeepMind's own post says official forecasts and warnings should come from your local meteorological agency or national weather service. The model is an input for professional forecasters, not a replacement for them.
What this means for businesses
Most small and medium businesses will never run a weather model, and should not try. The benefit arrives indirectly, through better official forecasts and through the commercial weather, insurance and freight services that build on open models like this one.
What a business can do is make sure an extra day of warning is actually used. A builder with a crane on site, a transport operator with trucks on a coastal highway, a cafe group with cold rooms full of stock and a wholesaler waiting on a container all make better decisions with 72 hours of notice than with 48. But only if someone has decided in advance what to do with it.
Trade is where the effect is most direct. Growers, packers and shippers work to harvest windows and vessel schedules that a single storm can upset. Truly Ceylon, a Sri Lankan export business that supplies spices, coconut products and other foods to importers and wholesalers, is one example of a business that depends on farms, roads and ports staying open through storm season. Australian importers on the other end of those routes carry the same risk in their lead times. Businesses in import and export and in logistics and transport have the most to gain from building weather into their planning.
A short checklist:
- Know your official source for warnings and who in the business watches it during storm season.
- List the operations that weather can stop: deliveries, site work, events, cold storage, staff travel.
- Agree trigger points in advance. For example, what happens when a cyclone watch is issued, and what happens when it becomes a warning.
- Check that business data is backed up somewhere outside the affected region and that staff can work from another location.
- Ask freight and supply partners how they will tell you about delays, and how early.
- Do not base safety decisions on experimental models or on apps that do not say where their forecasts come from.
If weather regularly disrupts your operations, it can help to put forecast and warning feeds next to your own order, job and delivery data so the people making decisions see both in one place. That is ordinary reporting work of the kind described on our data and dashboards page. If you would like to talk it through, you can send us an enquiry.
Key takeaways
- WeatherNext Cyclones was published in Nature on 6 August 2026 by Google DeepMind and Google Research.
- It forecasts cyclone track, intensity and size in one model and reports an average lead-time gain of one day or more over leading operational systems.
- It can produce up to 1,000 scenarios per forecast, and a 15-day forecast takes under a minute.
- The code and weights are public, but the models are experimental and are not endorsed by any government weather agency.
- For businesses, the value is in having a plan that uses the extra warning time.
Frequently asked questions
What is WeatherNext Cyclones?
It is an AI weather model from Google DeepMind and Google Research that produces ensemble forecasts of tropical cyclones: where the storm will travel, how intense it will be and how far its winds will extend. It generates scenarios up to 15 days ahead and was described in a Nature paper published on 6 August 2026.
How much earlier can it forecast a cyclone?
The paper reports an average lead-time advantage of one day or more over leading operational models, based on cyclones from 2023 to 2025. DeepMind describes this as three-day forecasts that are about as accurate as earlier two-day forecasts.
Is WeatherNext Cyclones free to use?
The code and model weights have been published on GitHub, with the code under the Apache 2.0 licence. The full models need high-end accelerator hardware, while the smaller mini models can run in a free Colab notebook. The repository describes all of them as experimental research rather than a supported product.
Does this replace official cyclone warnings?
No. DeepMind states that official forecasts and warnings should come from your local meteorological agency or national weather service. The model is a tool for forecasters, and its repository says it is not endorsed by any government meteorological agency.
How can a small business benefit from better cyclone forecasts?
Mainly by planning. Decide in advance what the business does when a watch or warning is issued, who makes the call, how stock, sites and vehicles are secured, and how staff and customers are told. Better forecasts only help if there is a plan ready to use the extra time.