This article reflects information available as of August 2026. WeatherNext remains a research and experimental model, and official disaster-prevention information should still come from national meteorological agencies.

In August 2026, Google DeepMind and Google Research published their WeatherNext Cyclones research in Nature and simultaneously released the code and pretrained weights for WeatherNext 2, WeatherNext Cyclones, and a lightweight version of the model. The easiest number to understand from the results is that forecasts of tropical cyclone track, intensity, and wind radii achieved an average lead-time advantage of more than 24 hours. In practical terms, a three-day WeatherNext Cyclones forecast can reach roughly the same level of accuracy that major operational models previously achieved only at a two-day lead time. The paper describes the scale of this improvement as approaching the cumulative progress made in operational tropical cyclone forecasting over the past decade.

This news looks very different from the AI agent security stories that have dominated the week, but it offers another useful way to think about improving model capabilities. Greater model capability does not have a fixed positive or negative meaning on its own. What matters is the system in which the model is deployed, the problem it is solving, and who ultimately acts on its output. WeatherNext forecasts are still not official warnings and do not replace meteorologists. But if the model can surface risks such as rapid intensification, track shifts, and expanding wind fields earlier within an ensemble forecast, professional forecasters gain additional time to interpret those possibilities.

What Did WeatherNext Cyclones Actually Improve? One AI Model Predicts Track, Intensity, and Wind Fields Together

Tropical Cyclone Forecasting Has Long Faced One Problem: Track and Intensity Operate at Different Scales

Tropical cyclone forecasting has long involved a difficult tradeoff. Where a storm travels is largely controlled by large-scale global atmospheric circulation, making global models well suited to track forecasting. But how much a storm intensifies depends on convection, thermodynamics, and other small-scale processes near the storm core, which usually require much higher-resolution specialized models. This has often created a familiar split: global models perform well on track, while high-resolution regional models are better at intensity, making it difficult for a single system to handle both equally well.

WeatherNext Cyclones does not solve this by training a separate small model focused only on the cyclone center. Instead, a single AI model learns both global weather evolution and historical tropical cyclone data, then predicts storm track, maximum wind speed, and wind-field structure together. Evaluations in the Nature paper covering tropical cyclones from 2023 through 2025 show that WeatherNext Cyclones achieved an average forecasting skill advantage of about a day or more across all three categories compared with major operational models. Google DeepMind explains this more intuitively: the accuracy of its three-day forecasts is roughly equivalent to what previous models achieved at two days.

Predicting Intensity at 28-Kilometer Resolution Was One of the Most Unexpected Results

Another counterintuitive aspect of the research is that WeatherNext Cyclones did not achieve these results by relying on extremely high resolution. The full model uses global data at approximately 0.25-degree resolution, or around 28 to 30 kilometers. Google DeepMind says this is roughly 100 times coarser than the high-resolution models traditionally used for tropical cyclone intensity forecasting. The lightweight WeatherNext Cyclones Mini uses a 1-degree grid, or about 111 kilometers. Although its performance is lower than the full model, it can still produce global weather and tropical cyclone forecasts.

The result does not mean resolution is no longer important. A more accurate interpretation is that the paper shows very high resolution may not be strictly necessary to achieve current state-of-the-art intensity forecasting skill. Coarser global atmospheric data may contain more information related to cyclone intensity than previously assumed. Google DeepMind itself acknowledges that researchers do not yet fully understand why the model retains so much predictive capability at this resolution, making it an open question for future research.

WeatherNext Was Trained on Nearly 20 TB of Weather Data and Almost 5,000 Historical Storms

WeatherNext Cyclones was trained on two kinds of data: long-term global atmospheric states and curated historical tropical cyclone observations. Google DeepMind says the end-to-end training used nearly 20 TB of global atmospheric data together with NOAA’s IBTrACS global tropical cyclone database, which contains nearly 5,000 historical storms. This allows the model to learn not only what happens around an individual cyclone center, but how storms form, move, and intensify within the broader global weather environment.

The model uses Functional Generative Networks, or FGNs, as its generative architecture. The goal is not to produce only one “most likely track,” but to efficiently generate many possible scenarios. During operational runs in 2025, the system produced 50 ensemble members at a time, similar in size to major global physics-based ensemble systems. By 2026, the system had expanded to 1,000 members per run. The value of a larger ensemble is not simply that the map shows more lines. It increases the chance that low-probability but high-impact outcomes in the tails of the probability distribution remain visible, including scenarios such as rapid tropical cyclone intensification.

Has WeatherNext Actually Been Used in Hurricane Forecasting? Hurricane Melissa in 2025 Was Already a Real-World Case

The U.S. National Hurricane Center Received WeatherNext Forecasts During the 2025 Hurricane Season

WeatherNext Cyclones was not first tested on real storms only after the paper was completed. The model documentation released on GitHub confirms that WeatherNextCyclones_<2025 was the version actually run during the 2025 Atlantic hurricane season. It was publicly known at the time as FNV3, while the version post-processed by the U.S. National Hurricane Center was called GDMI. Coauthors of the Nature paper also include researchers and forecasters from NOAA／NHC, CIRA, and the UK Met Office, meaning the research was not conducted entirely outside operational forecasting environments.

Hurricane Melissa in 2025 is one of the clearest examples. When Melissa was still a relatively weak tropical system, WeatherNext was already forecasting that it would rapidly intensify and make landfall in Jamaica at Category 5 strength. Five days in advance, the model assigned roughly an 80% probability to that outcome, rising to nearly 100% three days in advance. The NHC also, for the first time, forecast that a storm still producing only Category 1 winds would later reach Category 5 intensity. Google DeepMind says WeatherNext was one of several forecast sources used by NHC at the time, alongside physics-based models such as HAFS, satellite observations, hurricane reconnaissance aircraft, and professional forecaster judgment. The AI did not independently issue the warning.

This distinction matters. It is supported to say that “NHC used WeatherNext as part of its forecast guidance for Melissa.” Saying that “AI independently predicted the storm five days early and therefore caused officials to order evacuations” would stretch the causal claim too far. WeatherNext provided additional forecast guidance, while official warnings, risk assessments, and evacuation decisions remained the responsibility of meteorological and government agencies.

What Has WeatherNext Open-Sourced? Code, Model Weights, and the Mini Version Are All Available

Code and Pretrained Weights for WeatherNext 2 and WeatherNext Cyclones Have Been Released

Google DeepMind released more than the Nature paper. It also published a WeatherNext repository that can be used to run the models. GitHub currently provides code and multiple sets of pretrained weights for WeatherNext 2, WeatherNext Cyclones, and WeatherNext Cyclones Mini, including checkpoints used during the 2025 hurricane season as well as versions preserved to reproduce the 2023 and 2024 research results. WeatherNext 2 and WeatherNext Cyclones use the same cyclone forecasting approach. One major difference is that WeatherNext 2 additionally predicts 100-meter wind speed, while the two models use different weights because they were trained independently.

One licensing detail in the original discussion needs to be handled carefully: it is not accurate to say that “all code and model weights are licensed under Apache 2.0.” The GitHub documentation currently places Colab notebooks and related code under the Apache 2.0 license, while other released materials use CC BY 4.0. Pretrained weights are downloaded separately from a Google Cloud Bucket. A more accurate description is that WeatherNext’s research code and model materials have been released under open licenses, but different components are covered by different terms.

WeatherNext Cyclones Mini Can Run on a Single Free Colab TPU

The hardware requirements for the full WeatherNext models are still substantial. GitHub documentation states that non-Mini models require GPU memory at the level of an H100 when using GPUs. Mini was specifically designed to reduce memory and compute requirements and can run on a single GPU or a single TPU. Google’s provided Colab Notebook loads WeatherNext Cyclones Mini by default and recommends the v5e-1 TPU runtime currently available for free in Colab.

That makes “accessible to ordinary researchers” more accurate than “anyone can run the full WeatherNext model.” The complete model still requires significant hardware resources, but the Mini version substantially lowers the barrier for testing, education, and research. Research teams can first inspect data pipelines, trackers, and model outputs in a cheaper environment before deciding whether to invest in the compute required for the full version.

Why Does WeatherNext’s “Open Model” Story Look So Different from Other Recent AI Security News?

In several recent AI security stories, open weights have been discussed mainly as a control problem: once a model is downloaded, the original developer can no longer uniformly enforce API restrictions, suspend accounts, or monitor every deployed copy. WeatherNext follows a similar distribution logic. Once the code and model materials are public, research teams around the world can run them independently, modify research workflows, and potentially develop versions better suited to specific regional forecasting needs. The difference is not that “openness” suddenly becomes inherently safe, but what capabilities are being opened and what they are being used for.

For global weather forecasting, this kind of openness has a direct research benefit. Weather conditions are highly regional. Typhoons, monsoons, mountain rainfall, and coastal wind fields all present different forecasting needs. If national meteorological services, universities, and nonprofit research groups can begin with an already trained model, they do not need to first reproduce the enormous cost of training a global model from scratch. Google DeepMind explicitly lists specialized and localized models as one of the uses it hopes the released system will support.

However, openness does not mean WeatherNext output automatically becomes an official weather service. Google clearly labels the repository as an experimental research project. The model has not received official endorsement from any government meteorological agency and cannot replace formal alerts, warnings, or notices. Even though the Nature paper includes authors from NHC, CIRA, and the UK Met Office, responsibility for public warnings remains with national meteorological agencies.

Can Ordinary People Use WeatherNext Now? Weather Lab Already Lets Users View AI Weather Forecasts

You do not need to download the model to see experimental WeatherNext forecasts. Google DeepMind has updated Weather Lab so that, in addition to tropical cyclone tracks, users can now view WeatherNext forecasts for global temperature, precipitation, wind speed, and other variables. The WeatherNext repository also provides ways to access daily model output through Google Cloud, Weather Lab, and Open-Meteo, meaning researchers and developers do not necessarily need to run the entire model themselves to obtain forecast data.

Weather Lab is better suited for comparing how different models understand future weather than for replacing Taiwan’s Central Weather Administration or another official agency. Google DeepMind repeatedly emphasizes in both its WeatherNext announcement and GitHub documentation that official weather alerts should still come from the relevant national meteorological service. In Taiwan, typhoon warnings, work and school closures, and disaster-prevention decisions must still be based on official information from the Central Weather Administration and other government agencies. A WeatherNext track appearing farther from Taiwan is not sufficient reason to conclude that the risk has passed.

What WeatherNext Actually Means for Ordinary People Is Not “You Can Now Know Where a Typhoon Will Go One Day Earlier”

“One More Day” Means Better Forecast Skill, Not That Every Warning Will Be Issued 24 Hours Earlier

The day or more of lead-time advantage reported in the Nature paper is an average result derived from comparing errors across many tropical cyclone track, intensity, and wind-radius forecasts. It is not a promise that every future typhoon warning will always be issued exactly one day earlier, nor does it mean every five-day track suddenly becomes certain. WeatherNext still uses ensemble forecasting to represent multiple possible scenarios. The important improvement is that forecasters may be able to see certain risks accumulating earlier, not that they receive a perfectly certain answer sooner.

For disaster preparedness, that additional forecasting skill can still have real value. In rapid-intensification cases such as Melissa, if an ensemble begins assigning a higher probability to extreme intensification earlier, meteorological agencies have more time to incorporate that possibility into their assessments. Evacuations, resource allocation, port operations, and decisions over work or school closures still need to be made by local governments using the full range of meteorological information. WeatherNext is better understood as adding an earlier signal to those decisions.

After the Research Release, the More Important Question May Be How Different Regions Adapt the Model

After this release, the next stage is not only waiting for Google to improve another benchmark. It is also watching how research institutions in different regions use the system. Google DeepMind is already working with the U.S. NHC and has also publicly mentioned research collaborations involving the Philippines’ PAGASA, Taiwan’s Central Weather Administration, Indonesia’s BMKG, and Vietnam’s VNMHA. This does not mean WeatherNext has become the official forecasting system of these agencies. It means the model is entering more regional research and validation environments.

That is where an open model can have more practical value in meteorological research. A global model can provide a shared foundation, while researchers who genuinely understand local weather can test which situations the model handles reliably and where it tends to fail. For ordinary users, the final benefit may not be seeing the name of an AI model inside a weather app. It may instead be that official forecasts begin identifying rapid intensification or track risks earlier, with the model operating behind the scenes as part of a larger forecasting toolkit.

Many AI stories this week have focused on what happens when models become too capable at taking actions: how permissions should be restricted and how sandboxes should be secured. WeatherNext provides a different case. Model capabilities are advancing rapidly here as well, and the model materials are also being released openly, but the problem becomes how more meteorological agencies and research teams can validate, modify, and use them. That does not mean WeatherNext has no risks, nor does it need to be framed as a simplistic story of “AI finally being used for something good.” The more concrete development is that tropical cyclone track, intensity, and wind fields—three forecasting challenges that have historically been difficult to improve together—are now being handled by an AI system that achieved an average lead-time advantage of more than a day in evaluations covering 2023 through 2025. The next useful question is whether that advantage can remain stable across different ocean basins, different storm types, and real operational forecasting environments.

FAQ

SUPPORT FENGNIII

喜歡這篇文章嗎？

如果這篇內容對你有幫助，可以透過小額贊助支持本站持續整理更多日文、韓文、旅行與數位工具內容。

小額支持本站

付款將由藍新金流安全處理