Google DeepMind just posted a hard number to a soft problem: its new WeatherNext model gained more than a full day of lead time on tropical cyclones. According to Google DeepMind, that jump equals roughly a decade of traditional meteorological progress, delivered in a single research cycle. The team published the work in Nature and, in the same move, open sourced the code and model weights for anyone to use.
That combination, a genuine forecasting gain plus a full open release, is what makes this worth your attention.
What actually happened
WeatherNext uses something called Functional Generative Networks (FGNs) to produce large ensembles of forecasts. An ensemble is just many slightly different predictions run together, which is how forecasters measure uncertainty instead of betting on one outcome.
Here’s the scale shift, in Google DeepMind’s own numbers:
- Last year the system ran 50 predictions at a time, matching global physics models.
- This year it scaled to 1,000 ensemble members.
- A single 15-day forecast now runs in under a minute on one TPU.
More members means the model catches rare but brutal scenarios, like the rapid intensification seen during Hurricane Melissa in 2025. Those tail risks are exactly the events that catch communities off guard, so surfacing them faster has real stakes.
The part that surprised the scientists
For years the field assumed you needed very high spatial resolution to nail storm intensity. WeatherNext Cyclones breaks that assumption. It runs on data at 28x28km resolution, roughly 100 times coarser than traditional models, and still delivers accurate intensity forecasts.
A smaller version, WeatherNext 2-mini, runs at an even coarser 111x111km and still performs well. Google DeepMind is candid that it doesn’t fully understand why this works yet. The company calls it an open research question and is inviting the wider community to help figure it out.
That honesty matters. It’s a reminder that these systems are outrunning the theory that explains them, and the people who built this one are saying so out loud.
What’s being released
Google DeepMind is putting several things in the open:
- WeatherNext Cyclones: the model that ran during the 2025 hurricane season, with results in the paper.
- WeatherNext 2: a later update the team operationalized in October.
- WeatherNext 2-mini: a compact model that runs on a single TPU inside a free public Colab notebook.
- Weather Lab: a refreshed interface that now shows global forecasts for temperature, precipitation, and wind speed alongside cyclone tracks in one view.
All of it sits under the Google Earth AI umbrella.
Why it matters for the industry
The status quo in operational forecasting has been physics-based models running on supercomputers, expensive to operate and slow to iterate. AI weather models have been chipping away at that for a couple of years, but they were mostly closed or research-only.
This release changes the access equation. A meteorological agency, a university lab, or a nonprofit can now take the weights and build localized forecasting tools without training a foundation model from scratch. Running a 15-day forecast in under a minute on one TPU puts serious capability within reach of teams that could never afford a physics supercomputer.
What stands out here is the pairing of a measurable safety benefit with open weights. A full extra day of warning on a rapidly intensifying storm is the difference between an ordered evacuation and a scramble.
What to expect next
Google DeepMind is framing this as collaborative, not a replacement for human forecasters. The company explicitly points people back to their local meteorological agency for official warnings, and positions WeatherNext as a tool that pairs machine learning with human expertise.
A few things to watch:
- Whether national weather services fold these models into operational pipelines for the next storm season.
- What specialized, localized models the research community builds on top of the open weights.
- Progress on the open question of why coarse-resolution models forecast intensity so well.
If you work anywhere near climate risk, insurance, logistics, or emergency planning, this is worth tracking closely. Full details, the Nature paper, and live cyclone forecasts are available at the original source.