WeatherNext Is Now Open Source. Who in the Caribbean Will Test It?

Hurricane Melissa came ashore near New Hope, in Westmoreland, Jamaica, on 28 October 2025 with sustained winds of 185 mph and a central pressure of 897 millibars, tied with Hurricane Dorian and the 1935 Labor Day hurricane as the strongest Atlantic landfall by wind speed, according to the National Hurricane Center's final report. Google has since released the code and weights for an AI model that helped forecast that landfall. A Caribbean team can now evaluate its own copy, but the release alone does not establish whether doing so would improve local warnings.

TLDR

• WeatherNext Cyclones was published in Nature on 6 August 2026. Google released code and model weights, including a smaller model for demonstrations.
• The model helped NHC forecasters assess Melissa. Its strong track results did not remove intensity errors or the need for human judgement.
• Jamaica's estimated US$8.8 billion in physical damage and later US$12.2 billion in total damage and losses measure different things.
• A regional evaluation needs staff, suitable computing, observations and a clear comparison with existing guidance. Official warnings remain the responsibility of meteorological authorities.

What Did the AI Model Get Right About Melissa?

In its May 2026 account of Melissa, Google says its ensemble predicted a Category 5 landfall in Jamaica with about 80 percent confidence five days ahead and nearly 100 percent three days ahead. These are Google's retrospective figures for that forecast, not a general accuracy rate. The NHC also used physics-based models and observations; Jamaica benefited from the resulting guidance without needing to run the AI model itself.

Brightband's separate analysis found DeepMind's ensemble mean had the lowest track errors across the lead times it compared for Melissa. In the 23 October 2025 06 UTC cycle, DeepMind and HAFS anticipated the slow movement and turn towards Jamaica more successfully than GFS. DeepMind still brought landfall forward too quickly before later forecasts corrected it. This was one storm, and the comparison mixed an ensemble mean with individual dynamical models.

A better track forecast can help officials plan supplies and evacuations. Those decisions also depend on surge, rainfall, local exposure and advice from the national meteorological service. No single model determines which parish should evacuate.

What Did It Miss?

Brightband found that DeepMind had the lowest intensity errors at two to four days overall. Yet in the 23 October 06 UTC cycle, none of the compared models fully captured the following two to three days of rapid strengthening. The ensemble mean levelled off too weak. The full ensemble was publicly available, and forecasters did inspect individual members showing much deeper pressures. An average can hide an important tail risk.

The later Nature paper, published on 6 August 2026, evaluates WeatherNext Cyclones across 2023–2025 storms. Its broader performance results are stronger evidence than a single successful landfall forecast, but they still do not establish skill for every island, storm or local hazard. A local evaluation should report track and intensity errors separately.

Track and intensity forecasts support short-term preparations, including where to position supplies and which shelters to open. They do not set the structural standard for a roof, shelter or clinic. Building codes and engineering assessments must account for long-term local hazards, rather than the category predicted for one approaching storm.

What Did Melissa Cost, and Why Does the Forecast Matter?

The UN in Jamaica cites an early World Bank estimate of US$8.8 billion in physical damage. A later assessment put total damage and losses at US$12.2 billion, equivalent to 56.7 percent of 2024 GDP, as the IDB reported in April 2026. Physical damage and total damage and losses are different measures.

Earlier, reliable warning can give officials more time to move people and equipment. How much value an additional forecast adds depends on the decisions it changes, the warnings already available and whether communities can act. These damage estimates do not tell us the financial return from running a local copy of WeatherNext.

What Does Open Access Actually Provide?

The official repository provides code and notebooks under Apache 2.0 and other released materials under CC BY 4.0. Third-party datasets have their own terms. Its demonstration defaults to WeatherNext Cyclones Mini on a free Colab v5e TPU; the larger models require different hardware. The repository warns that Mini is not expected to match the larger models' performance.

DeepMind's release announcement describes a single 15-day forecast taking under a minute on a TPU and ensembles of up to 1,000 scenarios. Those figures do not mean a complete local forecasting service is free. A team still needs observations, staff, computing capacity, version control and a plan to validate and maintain the system.

Who Could Help Evaluate It?

The NHC provides regional tropical cyclone guidance, while national meteorological services issue local advice and warnings. In Barbados, the Caribbean Institute for Meteorology and Hydrology supports training, research and climate services for sixteen Caribbean Meteorological Organization governments. Its predecessor meteorological institute began in 1967; the current name followed a later merger. CIMH offers a meteorology degree with UWI Cave Hill and could be a partner in a regional test, alongside national services. That is a proposal, not an assigned mandate.

Local evaluation need not duplicate everything the NHC already does. A useful question is whether running a regional copy adds information that existing guidance does not provide, and whether that benefit justifies the work. The open release does not tell us which Caribbean agencies already use its outputs or are testing it internally.

Our Affiliation

CAIA is part of the same network as StarApple AI, founded by CAIA president Adrian Dunkley. His own site describes climate-related AI research. Any local proposal, including one from this network, should face the same independent comparisons with observations and operational baselines. An affiliation or a physics-informed design does not establish forecast accuracy.

How Should Residents Read an AI Forecast?

Follow official watches and warnings from your national meteorological service and the NHC. An experimental forecast is not a replacement. If someone shares an AI run, ask which model and version produced it, when it was initialised, and whether the number represents one ensemble member or an average. A striking scenario is not a probability assessment on its own.

Google's high confidence in a Category 5 landfall and Brightband's intensity criticism are not mutually exclusive. Predicting the category and broad landfall does not mean predicting the exact peak strength or the full rate of intensification. The results also refer to particular forecast cycles and summaries of the ensemble.

What Would a Caribbean Test of the Model Look Like?

A useful first test could be a retrospective comparison. Take the open model and run it on the seasons the region already remembers: Beryl in 2024, which the UN in Jamaica names alongside Melissa as back-to-back hurricanes, and Melissa in 2025. Compare each run with what local stations recorded, publish where it beat the operational guidance and where it did not, and hand the results to the meteorological services in the countries that must act on the warnings. The exercise needs staff time, computing and suitable observations. It would produce a local error record that forecasters could inspect. Use model checkpoints trained before the test storms to avoid evaluating on training data.

What Should Happen Next?

Before launching a test, participating meteorological services and researchers should agree on who owns it, the staff and computing budget, and which questions the results must answer. Record the model version, training cutoff, forecast cycle and comparison data. Publish errors and misses alongside successes, then let the relevant authorities decide whether any operational use is justified.

Universities could use the smaller demonstration model for teaching while a properly resourced team evaluates the larger system. Those are different tasks. Neither should be presented to residents as a new official warning service.

CAIA's review of Melissa's drone response also examines practical disaster-response capacity. Melissa's first anniversary falls on 28 October 2026. A published evaluation plan would be a useful step towards deciding how the region can use this release.

Supporting resources: StarApple AI; Adrian Dunkley; Caribbean AI Risk Management Council; AI Jamaica; AI Barbados; 14West. Related reading: the Caribbean AI Forum report.