WeatherNext Gained a Day on Cyclone Forecasting. Open Code Will Show Where It Holds

Google's new cyclone model reached a striking benchmark result: averaged across its evaluation, a forecast made three days before a storm was about as accurate as forecasts earlier leading systems produced two days before it. The comparison covered the cyclone's route, strength, and surrounding wind field.

That extra day is the headline from Google DeepMind's WeatherNext announcement. The more consequential development may be what followed. Google released the code, pretrained checkpoints, documentation, and a public notebook, giving researchers outside the company a way to examine where the advantage holds and where it does not.

The benchmark does not mean every coastal community will automatically receive another 24 hours to evacuate. It means the relationship between forecast accuracy and lead time moved in a useful direction. Converting that shift into an official warning still depends on timely observations, robust data pipelines, regional verification, human interpretation, and the responsible meteorological agency.

WeatherNext Cyclones has produced enough evidence to deserve close attention. Its open release now lets the weather community test how much confidence the system deserves.

The extra day is an accuracy comparison, not a universal promise

The Nature study on operational tropical cyclone forecasting with AI evaluates the model against historical cyclones from 2023 and 2024. According to Google, the comparison included both single deterministic predictions and distributions of possible outcomes from leading global and cyclone-focused systems.

Across track, intensity, and wind structure, WeatherNext delivered an average lead-time improvement exceeding 24 hours. Put plainly, its three-day output reached roughly the accuracy that previous models achieved at two days.

Each part of that result matters for a different operational question. Track estimates where the cyclone could travel. Intensity describes how powerful it could become. Wind structure maps how damaging winds may extend beyond the center. A precise center line is not enough if the system misses how quickly the storm is strengthening or how widely its winds spread.

The word average keeps the result in bounds. The evaluation combines different storms, forecast horizons, regions, and performance measures. It does not establish that every WeatherNext prediction improves by one day. It also cannot guarantee equal performance on a storm that differs sharply from the examples represented in the model's training and evaluation data.

The next useful questions are more granular. Researchers need to determine which basins and storm types account for the gain, where it becomes smaller, and whether the probabilities attached to different outcomes agree with what subsequently happens.

One model is trying to resolve two forecasting scales

Cyclone prediction contains a stubborn split. Large atmospheric patterns steer a storm over long distances, which makes global models valuable for estimating track. Changes in intensity depend heavily on smaller processes around the cyclone's core, where specialized models using finer resolution have traditionally held an advantage.

WeatherNext Cyclones attempts to learn both scales within one system. Google reports that its end-to-end training used almost 20 terabytes of atmospheric information together with expert-curated entries from the International Best Track Archive for Climate Stewardship. Those records cover close to 5,000 historical storms. The resulting model predicts global weather while producing cyclone-specific forecasts for path, intensity, and wind structure.

Its spatial resolution makes the result especially interesting. The main model consumes inputs on a grid of roughly 28 by 28 kilometres. DeepMind says this is around 100 times coarser than the traditional input resolution commonly associated with accurate intensity prediction. A smaller version operates at approximately 111 by 111 kilometres.

The development team says it does not yet completely understand how WeatherNext retains its cyclone skill at that level of detail. That is not a minor footnote. If a model appears to recover useful small-scale behaviour from coarse inputs, researchers need to discover what information it is using, when that mechanism works, and when it may fail.

Open access makes that uncertainty productive. Independent teams can investigate the mechanism instead of treating the benchmark as its own explanation.

Fast forecasts make larger ensembles possible

A useful cyclone forecast cannot consist only of one confident path. The atmosphere is uncertain, and small changes in its initial state can produce meaningfully different tracks or intensities several days later. Forecasters therefore inspect ensembles: multiple plausible futures generated from related starting conditions.

WeatherNext creates those alternatives using Functional Generative Networks. DeepMind reports that one 15-day forecast runs in less than a minute on a TPU. During the 2025 hurricane season, the team generated 50 scenarios for each cyclone. It says the ensemble has since grown to 1,000 scenarios.

Those runtime and ensemble figures need to remain separate. The reported sub-minute speed applies to one forecast, not an entire batch of 1,000 members. The reasonable connection is that inexpensive individual runs make much larger ensembles practical.

More scenarios can help expose low-probability outcomes such as rapid intensification. They also generate far more information for forecasters to assess. Their decision value depends on calibration. If the system assigns a class of events a 10 percent probability, comparable events should occur at approximately that frequency over a suitable evaluation set.

Speed removes a computational obstacle. It cannot decide whether a probability deserves trust, how a regional agency should interpret it, or which action is appropriate when different models disagree.

The repository turns the announcement into something researchers can test

The public WeatherNext repository now includes WeatherNext 2 and WeatherNext Cyclones code, pretrained checkpoints, technical documentation, and an interactive notebook. The available releases serve distinct purposes:

1. WeatherNext 2 provides the model at approximately 30-kilometre resolution, trained through 2024 and used operationally.
2. WeatherNext Cyclones supplies checkpoints connected to the 2023 and 2024 paper results, along with the version run during the 2025 Atlantic hurricane season.
3. WeatherNext Cyclones Mini offers 1-degree checkpoints intended for experiments with lower memory requirements.

Naming could cause some early confusion. DeepMind's announcement refers to the compact release as WeatherNext 2-mini, while the repository lists the downloadable versions as WeatherNext Cyclones Mini. The repository also cautions that these smaller checkpoints are not expected to reproduce the performance of the larger models.

That distinction matters because the easiest demonstration uses the smaller release. Google's public Colab notebook selects the Mini model by default and can run with a free v5e-1 TPU. According to the repository, larger variants need stronger accelerators, such as a v5p TPU or an H100 GPU after changing the attention implementation.

Anyone can inspect the notebook without first assembling Google's complete research environment. Reproducing the full-scale results still carries a meaningful compute requirement.

The licensing supports practical research reuse. Code and notebooks are released under Apache 2.0, while other published materials use CC BY 4.0. Pretrained weights and sample data are available from a Google Cloud bucket. Teams must separately review the conditions governing the ECMWF data involved in training or initialization.

This is the real benefit of the release. Researchers can compare results across ocean basins, inspect the cyclone tracker, evaluate probability calibration, reproduce the paper checkpoints, and test regional adaptations. Open sourcing does not answer any of those questions automatically. It allows people outside Google to produce evidence about them.

Hurricane Melissa shows how the model can enter a real workflow

WeatherNext had already been used as guidance during an active storm. In Google's account of Hurricane Melissa, the system forecast a Category 5 landfall in Jamaica five days in advance with 80 percent confidence. That confidence rose to almost 100 percent three days before landfall.

The National Hurricane Center did not treat the AI output as a replacement for its other evidence. Forecasters considered it alongside physics-based models, satellite measurements, and data collected by hurricane hunters.

That is a credible picture of how a consequential forecasting model should contribute. It offers another source of guidance while trained specialists compare it with observations and established systems. The case demonstrates that WeatherNext can be useful inside an expert workflow.

One exceptional cyclone cannot establish reliability across every basin, observation network, or rapid-intensification pattern. The historical evaluation provides broader evidence, and operational agencies still need local verification before relying on the model in their own warning processes.

The repository states the boundary directly. WeatherNext is experimental research software without promises of API stability, and it is not a substitute for official alerts. Its larger models also rely on operational initial conditions and supporting infrastructure. A checkpoint can be a valuable forecasting component without becoming a complete warning service.

Five tests now matter more than the launch headline

The open release moves WeatherNext from a company announcement into a form that can be challenged. Five questions should guide that work:

1. Reproduction: Can independent researchers recover the reported lead-time gain from the published checkpoints?
2. Variation: How does accuracy change across basins, cyclone strengths, forecast horizons, and different qualities of initial observations?
3. Calibration: Do the ensemble probabilities match observed frequencies, particularly for rare intensification and damaging wind outcomes?
4. Generalization: Does the 28-kilometre model preserve its intensity performance on storms that differ from its historical training record?
5. Disagreement: When WeatherNext and physics-based guidance diverge, can forecasters identify the conditions under which either system is more likely to fail?

These questions set a higher standard than asking whether the model won a benchmark. A forecasting system earns operational trust by showing where it is dependable, exposing where uncertainty grows, and remaining useful when the atmosphere does something unusual.

Open code starts the operational test

WeatherNext Cyclones has produced a result that would matter if it persists: comparable forecast accuracy one day earlier, on average, across the path, strength, and wind structure of a cyclone. That could give forecasters more time to compare plausible outcomes before a storm arrives.

The qualification is part of the result. Historical accuracy is not the same as a universal extra warning day. Fast ensembles are not automatically calibrated ensembles. A successful forecast during Hurricane Melissa is evidence, but it cannot stand in for verification across regions and storm types.

Google's open release makes those boundaries inspectable. Researchers can now test the checkpoints, divide the aggregate result into harder regional questions, and examine whether larger ensembles improve decisions without disguising uncertainty.

The benchmark earned WeatherNext serious attention. The next measure of progress is whether independent evaluation and operational use can turn that measured day of predictive skill into guidance that forecasters know when to trust.