DeepMind Just Open-Sourced a Better Hurricane Forecast

For thirty years, hurricane forecasting had a dirty secret: we got steadily better at predicting where a storm would go and barely better at predicting how strong it would get. Track errors shrank decade after decade as global models improved. Intensity forecasts flatlined, because rapid intensification lives in the storm's inner core — convection happening at scales far below what a global model's grid can resolve. NOAA's answer was to bolt specialized regional models like HWRF onto the global ones, trading worldwide coverage for local detail.

On August 6, Google DeepMind published a Nature paper claiming its WeatherNext cyclone model breaks that trade-off — and, more interestingly for anyone who builds software, released the whole thing: code, weights, and ensemble forecast feeds you can query today.

What actually got better

The headline number is a day. DeepMind's three-day cyclone forecasts match the accuracy that the best operational systems previously delivered at two days, across track, intensity, and wind structure. Given that forecast skill historically improves by roughly a day per decade, that's a decade of progress in one release — and it's the intensity result that matters most. A 28-kilometer-resolution global model beating purpose-built regional models like HWRF at predicting intensification was, until recently, close to heresy in operational meteorology.

The validation isn't just benchmarks. The U.S. National Hurricane Center ran WeatherNext alongside its operational guidance through the 2025 Atlantic season. When Hurricane Melissa was still a disorganized Category 1, DeepMind says the model put 80% probability on it reaching Category 5 — five days before its catastrophic Jamaica landfall. Forecasters got the rapid-intensification call right with lead time they've never had before. That's an operational agency staking real warnings on ML output, which is a far higher bar than a WeatherBench leaderboard.

Under the hood this is the payoff of a line of work that's been compounding fast: GraphCast showed in 2023 that a learned model could beat physics-based deterministic forecasts; GenCast made it probabilistic in 2024; ECMWF put its own AIFS model into operations in early 2025. WeatherNext Cyclones uses Functional Generative Networks — noise injected into the architecture itself — to spin up ensembles of 1,000 scenarios per cyclone, up from the 50 that was standard a year ago. Each 15-day forecast takes under a minute on a single TPU. A physics-based ensemble of that size would need a supercomputer hall and hours of wall-clock time. Training took roughly 20 TB of atmospheric reanalysis plus the IBTrACS archive of about 5,000 historical storms.

The part developers should care about

The release comes in three consumable forms, and picking the right one matters.

Hosted data, no ML required. WeatherNext 2 forecasts — all 64 ensemble members, refreshed every six hours — land in BigQuery and the Earth Engine data catalog, with Vertex AI hookups if you want to build downstream models. If you're doing logistics routing, parametric insurance triggers, or energy-demand forecasting, this is the path: a SQL query against a public dataset instead of a NOAA GRIB-file ingestion pipeline, which anyone who has parsed GRIB2 will recognize as a genuine quality-of-life upgrade. Open-Meteo also serves the model through its API, which gets you WeatherNext output in a REST call with no GCP account at all.

Self-hosted weights. The GitHub repo ships WeatherNext 2, three vintages of WeatherNext Cyclones (including the exact 2025 model the NHC ran live), and 1°-resolution mini variants — Apache 2.0 code, CC BY 4.0 weights. The mini models run in a free Colab TPU runtime, which makes for a legitimately great afternoon of tinkering. The full models want a v5p TPU or an H100, and here's the catch: they initialize from ECMWF HRES analysis fields. The model is free; the freshest initial conditions that feed it are not fully open. Your self-hosted forecast is only as current and as good as the analysis data you can obtain.

That catch is the tell for where this market is going. The moat in weather prediction used to be the model — decades of physics code and the supercomputers to run it. DeepMind just gave the model away. What's left as scarce inputs are data assimilation (turning raw satellite and sensor observations into the clean gridded state the model starts from) and observation networks themselves. National met agencies aren't being disintermediated; they're being repositioned as upstream data providers, with the forecast layer commoditizing above them.

Genuine shift, with two honest caveats

This is not hype. Peer review, a full operational season at the NHC, open weights, and public data feeds is about as complete an evidence package as an ML systems release ever gets. Proprietary weather-data vendors selling repackaged model output at enterprise prices should be nervous; so should anyone whose product's value was "we have forecast data and you don't."

Two things temper the enthusiasm. First, ML weather models learn from history, and a warming climate keeps manufacturing storms with weak precedent — Melissa-style rapid intensification is itself becoming more common. The 1,000-member ensembles are partly a hedge against exactly this tail risk, but out-of-distribution behavior remains the standing question for every learned forecaster. Second, probabilistic output demands probabilistic engineering. If your app grabs the ensemble mean and displays it as "the forecast," you've thrown away the entire point. The work is designing for percentiles — trigger the shipping-delay warning at the 90th-percentile wind speed, price the insurance product off the full distribution — and that's a product-design shift, not a data-source swap.

The realistic adoption path: prototype against Open-Meteo this week, move to BigQuery when you need ensemble members and SLAs, and skip self-hosting unless forecasting is your core product. The forecast just became an API. The interesting engineering now starts on the consuming side.