DeepMind Open-Sourced the Cyclone Model the NHC Already Used

TL;DR

On August 6, Google DeepMind published "Operational Tropical Cyclone Forecasting with AI" in Nature and pushed the model to GitHub. WeatherNext Cyclones predicts a storm's track, intensity, and wind structure about a full day earlier than the systems it competes with: a three-day forecast now lands at the accuracy older systems needed two days to reach. It reads a coarse 28 km grid, samples 1,000 possible futures per run, and finishes a 15-day global forecast in under a minute on a single TPU. The code is Apache 2.0, the weights are in a public bucket, and the small variant runs in a free Colab notebook.

What actually shipped

This is a weights drop, not a demo page. The repo carries three current models plus the lineage that got here:

• WeatherNext 2, the general medium-range global forecaster, at roughly a quarter-degree grid.
• WeatherNext Cyclones, the storm specialist that the Nature paper is about, predicting track, intensity, and wind structure from one model instead of the usual stack of separate ones.
• WeatherNext 2-mini, a one-degree variant that DeepMind says still performs well and is small enough to run in a free Colab notebook.

GraphCast and GenCast, the previous generations, are still in the same repository. Licensing is split the way Google usually splits it: the code and notebooks are Apache 2.0, the accompanying materials are CC BY 4.0, and the pretrained weights sit in a public Google Cloud bucket you can browse without an account. A hurricane model you can wget is a genuinely strange sentence to be typing in 2026.

A day is not a rounding error

The headline result is lead time, not a leaderboard score. Cyclone forecasting is judged by how early you can say something useful, because the expensive decisions (evacuation orders, rig shutdowns, moving aircraft off an island) are gated on confidence at a specific hour. DeepMind's claim is that its three-day forecast is as accurate as what earlier systems produced at two days, across track, intensity, and wind structure. That is roughly 24 extra hours of warning at the same skill level, which the team frames as about a decade of meteorological progress folded into one model.

The trick is sampling, not certainty

WeatherNext 2 is built on Functional Generative Networks, and the important consequence for anyone who works with models is that it does not emit one answer. It emits a distribution, and the system draws 1,000 members per forecast. You do not read the output as "the storm goes here." You read it as "in 340 of 1,000 rolls the eye crosses this coastline," which is exactly the shape of a decision an emergency manager can act on.

The intuition: a deterministic forecast is one very confident friend telling you where the storm will land. A 1,000-member generative ensemble is a thousand slightly different friends, each fed the same imperfect starting data, and you count how many of them point at your town. The second one is less satisfying and much more useful.

Traditional ensembles do this too, by rerunning a physics simulation with perturbed initial conditions. The difference is cost. Perturbed physics runs are the reason ensemble sizes have historically been counted in dozens. Sampling a learned generative model is cheap enough that a thousand members stops being a budget line.

Coarse in, fast out

The counterintuitive part is the input. WeatherNext Cyclones works on a 28 km by 28 km grid, which DeepMind describes as about a hundred times coarser than conventional models, and the mini variant drops to 111 km and still holds up. Hurricanes have structure measured in kilometers, so a naive reading says the model should be blind to the thing it is predicting. It is not, because it learned the relationship between the large-scale flow it can see and the storm-scale behavior it cannot, which is a lesson every practitioner who has fought over input resolution should sit with for a minute.

It already ran a real season

This is the part that separates the release from a benchmark press release. The model was used operationally during the 2025 Atlantic hurricane season, and the National Hurricane Center drew on it while forecasting Hurricane Melissa's rapid intensification and landfall in Jamaica. Rapid intensification is the failure mode operational meteorology has been worst at for decades, and it is the one that kills people, because a storm that jumps two categories overnight outruns the evacuation plan built for the forecast you had yesterday.

DeepMind also credits collaboration with the NHC, the UK Met Office, and other national agencies. That matters for a model like this: without an operational partner willing to run it live and score it against what actually happened, you have a paper.

What a builder should take from this

You are probably not forecasting hurricanes. Three things here still transfer.

Uncertainty is a product feature, not a caveat

The value of a generative forecaster is not a better point estimate, it is a usable distribution. If your model emits a single answer with a confidence score bolted on afterwards, you have a worse version of this. Sampling the model many times and reporting the spread is often the cheapest real improvement available.

Coarse inputs are underrated

Everyone's instinct is to feed the model more resolution. This one wins at a hundredth of the input detail because the learned representation carries the relationships that the fine grid was being used to approximate.

Downloadable beats accessible

A Nature result behind an API is a product. A Nature result with weights in a bucket is infrastructure. Research groups and national agencies without a TPU budget can now fine-tune this on their own basin, which is the entire point of publishing the weights and not just the numbers.

The caveats

• Weights are not a forecasting service. You still need to source and preprocess the initial-condition data these models eat, which is the unglamorous part and the part that will consume your weekend.
• One hurricane is an anecdote. The Melissa case is compelling and it is not the evaluation. The Nature paper is the record; skill varies by basin, by storm, and by what you are scoring.
• "A decade of progress" is the vendor's framing. It is a reasonable way to describe a 24-hour lead-time gain given how slowly that metric has moved, but it is DeepMind grading DeepMind's homework, and physics-based centers like ECMWF will publish their own comparisons.
• Read the license split before you ship. Apache 2.0 on the code does not automatically describe every artifact in the release.

Key Takeaways

• DeepMind published WeatherNext Cyclones in Nature on August 6 and released code and weights on GitHub, with the code under Apache 2.0.
• The headline gain is roughly 24 hours of lead time: three-day forecasts at the accuracy earlier systems reached at two days, for track, intensity, and wind structure.
• It runs on a 28 km grid (111 km for the mini variant) and produces a 15-day global forecast in under a minute on one TPU.
• Each forecast is 1,000 generative samples, not one deterministic run, which is what makes the output a probability instead of a guess.
• It was already used operationally in the 2025 Atlantic season, including NHC forecasting of Hurricane Melissa's rapid intensification.
• Three variants ship: WeatherNext 2, WeatherNext Cyclones, and WeatherNext 2-mini, the last of which runs in a free Colab notebook.

Sources: Google DeepMind blog, Nature: Operational Tropical Cyclone Forecasting with AI, google-deepmind/weathernext on GitHub, The Keyword, Engadget, NOAA IBTrACS