Google DeepMind published WeatherNext Cyclones in Nature and open-sourced the model weights

WeatherNext Cyclones is a cyclone forecasting model that Google DeepMind published in Nature on August 6, 2026, releasing the code and model weights alongside the paper. The model predicts cyclone track and intensity in a single system and gains more than a full day of lead time on average, an improvement DeepMind describes as roughly a decade's worth of meteorological progress measured against the last 20 years of trends. Across 2023 to 2025, three-day position error is around 100 km against ECMWF-ENS and intensity error is around 11 kt against HWRF. ASAP works from DeepMind's official blog and the Nature paper to lay out what the model changes and how far the evidence carries.

One model started predicting track and intensity at the same time

WeatherNext Cyclones is a 2026 Google DeepMind model designed to predict cyclone track and intensity together rather than splitting the task across two model families. DeepMind frames this design as resolving a trade-off that forecasting has carried for decades.

The old division of labor had a physical basis. Where a cyclone goes is steered by massive global atmospheric currents, and those currents were best modeled by coarser global models. How strong a cyclone gets is driven by highly localized, fine-scale thermodynamic processes around its core, and those were the domain of specialized higher-resolution local models. A single phenomenon was handled by two different classes of model because each class was better at one half of it.

WeatherNext Cyclones crosses that boundary three ways. It co-trains on two distinct data modalities, global weather dynamics and expert-curated historical cyclone observations. It uses Functional Generative Networks (FGNs) to efficiently produce ensembles of different predictions, capturing the inherent uncertainty of the weather. And it operates at a comparatively coarse 28x28 km resolution. Training used nearly 20 terabytes of global atmospheric data together with the IBTrACS historical database spanning nearly 5,000 historical storms.

What the 100 km and 11 kt figures are actually measured against

The roughly 100 km three-day position error is measured against ECMWF-ENS, the European ensemble forecast system. The roughly 11 kt intensity error over the same window is measured against HWRF, the Hurricane Weather Research and Forecasting model. The verification period runs from 2023 to 2025.

Naming both baselines matters because the two numbers face different opponents. Track is compared against a global ensemble physics system, intensity against a hurricane-specific high-resolution model. The claim is that one AI model beat the specialist in each half separately, which is why the "we merged the division of labor" framing and the performance claim reinforce each other. It is not a single composite score.

The lead-time figure deserves the source's exact wording as well. DeepMind reports gaining more than a full day of lead time on average for track, intensity, and wind structure, and calls that improvement roughly a decade's worth of meteorological progress according to trends over the last 20 years. Decade here is a conversion, not an absolute yardstick: the slower a field's historical improvement rate, the more years any given gain converts into. Reading the number that way keeps it accurate.

Translated into operations, 24 hours is a decision unit. Evacuation orders, port closures, and utility shutdowns take physical time to execute, and whether one more day exists often determines whether the action is possible at all. Cutting error from 100 km to 90 km and getting the same accuracy a day earlier are different kinds of gain on the ground.

The puzzle left behind by 28x28 km working at all

DeepMind states that WeatherNext Cyclones needs data at only 28x28 km resolution, 100x coarser than traditional models. A smaller version, WeatherNext 2-mini, shows strong performance at an even coarser 111x111 km.

DeepMind writes that this has surprised scientists and that fully understanding how the models produce such accurate predictions at this resolution remains an open research question. Publishing organizations rarely record in their own announcement that they cannot explain their result, which makes that sentence the most notable line in the release.

The physics assumption was explicit. Predicting intensity was understood to require resolving eyewall convection on the grid, which required a fine grid, which cost enormous compute. If intensity error is lower on a grid 100x coarser, then intensity as an outcome is substantially inferable from large-scale atmospheric state without directly simulating core microstructure. There is a regime where learning the relationship between large-scale state and physical outcome beats simulating the physics.

That gain carries a cost. A physics model that fails can be inspected term by term; a model whose mechanism is not yet explained offers less to point at when it fails. High average accuracy and predictable failure modes are separate assets, and in disaster forecasting, where cost concentrates in the tail, the second one matters more than usual.

What scaling from 50 to 1,000 scenarios means for a forecaster

DeepMind scaled ensemble size to 1,000 members this year, having produced 50 predictions at a time last year, matching global physics models. A single 15-day forecast now runs in under a minute on a TPU.

A 20x increase in ensemble size is not an investment in average accuracy. DeepMind's stated purpose is capturing rare but consequential scenarios such as rapid intensification events, as occurred during Hurricane Melissa in 2025. The choice encodes a judgment that the forecaster's real problem sits in the tail of the distribution, not at its center.

The arithmetic makes the difference concrete. With 50 samples, a scenario with 2% probability appears once on average and frequently not at all in any given run. With 1,000, the same scenario appears about 20 times, which is enough to estimate the probability itself. That is the difference between telling a decision-maker that rapid intensification is possible and telling them it is a specific percentage, and for asymmetric-cost decisions like evacuation, that difference changes the call.

The sub-minute TPU runtime is what makes this reachable. Running 1,000 ensemble members through a global physics model was never on the menu at realistic compute budgets. This is a case where a change in cost structure changed the shape of the forecast product, and it suggests the practical contribution of AI weather models lies less in a single accuracy figure than in how many times you can afford to run them.

What Hurricane Melissa proved and what it did not

During the 2025 hurricane season, DeepMind's model helped the National Hurricane Center make a historic forecast for Hurricane Melissa by predicting the storm's rapid intensification and landfall in Jamaica. That enabled the NHC to issue an advance warning, giving teams on the ground critical time to prepare.

The evidentiary weight of a single success should be handled carefully. One forecast turning out right does not establish general reliability; the 2023 to 2025 aggregate verification does that work. What the Melissa case actually demonstrates is an integration path rather than an accuracy level, showing that model output entered an official warning procedure and was used there.

The agency question also deserves the source's exact framing. DeepMind's wording is that the model helped the NHC, and the body that issued the warning was the NHC. DeepMind attaches a note directing readers to their local meteorological agency or national weather service for official forecasts and warnings. In this arrangement the AI model is an input to forecasters, while judgment and accountability stay with the institution.

What open weights mean outside the agencies that already forecast

DeepMind open-sourced the code and model weights for WeatherNext 2 and WeatherNext Cyclones alongside the Nature paper on August 6, 2026. WeatherNext 2-mini runs on a single TPU in a free public Colab notebook, and current cyclone forecasts are browsable on Weather Lab, which is part of Google Earth AI.

The weight of that release is mostly about entry cost. Operating a global physics model requires supercomputer-class infrastructure and an organization able to run it, and that threshold has kept operational forecasting inside a small set of national agencies. If weights are open and a reduced model runs in a free notebook, a university lab or a disaster-response organization can run its own ensembles for the first time as a realistic option.

A limit needs stating alongside it. The published verification is a global aggregate, and this release does not report regional breakdowns such as skill for landfall tracks in any specific basin. Leading on a global average does not guarantee the same margin in a particular sea and season, so anyone evaluating regional deployment has to verify separately.

The larger barrier is not technical. Weather warnings are public outputs with legal responsibility and administrative procedure attached, and inserting a new model into that chain requires settling accountability and operational process, not just accuracy validation. What the NHC demonstrated with Melissa was an added input supporting forecaster decisions, not a replacement of the operational model. Open weights do not translate directly into operational substitution.

Open questions

The first is regional performance. The published position and intensity errors are global aggregates from 2023 to 2025, and how skill varies by basin or season is not knowable from this release alone.

The second is mechanism. Why intensity prediction works this well at 28x28 km is left as an open research question by DeepMind itself, and until that is answered there is correspondingly little basis for predicting the conditions under which the model breaks.

The third is behavior outside the training distribution. Training covers nearly 5,000 historical storms, and how a model fit to that distribution behaves in intensity regimes with sparse historical observation is a separate validation problem.

Source: ASAP analysis based on the Google DeepMind blog post "WeatherNext: AI model achieves breakthrough in forecasting cyclones" (August 6, 2026) and the Nature paper published the same day