the short version
- WeatherNext Cyclones reaches state-of-the-art track, intensity and wind-structure accuracy on 28x28km input data, breaking the assumption that intensity forecasting requires very high spatial resolution.
- The ensemble is the product: DeepMind went from 50 members last year to 1,000 this year specifically to capture rare tail events like rapid intensification.
- A single 15-day forecast takes under a minute on a TPU, but the blog does not state the TPU generation or the cost of a full 1,000-member run.
- The reported baselines are ECMWF-ENS for 3-day track and HWRF for intensity, evaluated on 2023-2024 storms — reproduce against those before trusting the lead-time claim.
On August 6, 2026, DeepMind published a Nature paper and a blog post describing WeatherNext Cyclones, an AI model that predicts a tropical cyclone's track, intensity and wind structure in one system. It also announced that WeatherNext 2 and WeatherNext Cyclones — the models used during the hurricane season — are being open sourced. For anyone who has been waiting to build weather tooling on something other than a vendor API, that is the sentence that matters.
The headline number is lead time. DeepMind reports that on average the model gains more than a full 24 hours: three-day forecasts are as accurate as what prior models produced at two days, which the team characterises as roughly a decade of meteorological progress at the trend rate of the last 20 years. The evaluation was run on historical cyclones from 2023 to 2024, benchmarked both deterministically and probabilistically against other top weather models.
The operational claim is separate and worth reading carefully. During the 2025 hurricane season the model helped the National Hurricane Center forecast Hurricane Melissa, predicting rapid intensification and landfall in Jamaica in time for an advance warning. The work involved forecasters at the NHC, CIRA, the UK Met Office and other agencies. The blog does not state the licence attached to the open-sourced weights, and does not link a repository in the text available here.
The inputs are coarser than expected
The most consequential engineering detail is resolution. Cyclone intensity has traditionally been treated as a fine-scale thermodynamic problem near the storm core, which is why intensity has been handled by specialised high-resolution local models while track came from coarser global models. WeatherNext Cyclones collapses that split and does it on data at 28x28km.
WeatherNext Cyclones only needs data with a resolution of 28x28km, 100x coarser than traditional models.
A smaller variant, WeatherNext 2-mini, runs at 111x111km and also performs well. DeepMind says this surprised scientists and that why the models work at this resolution remains an open research question. Note that the 100x figure is not disaggregated in the post — whether that is a linear or area comparison changes what it means, and it is the kind of thing to pin down from the paper rather than the blog.
The training setup explains part of the result. The model was co-trained on two data modalities: global weather dynamics and expert-curated historical cyclone observations, end-to-end across nearly 20 terabytes of global atmospheric data plus the IBTrACS database covering nearly 5,000 historical storms. The cyclone-specific supervision is doing work that raw resolution was previously assumed to be doing.
- Coarse inputs mean the ingest side of a deployment is far cheaper than a high-resolution physics pipeline — the bottleneck moves to sampling, not to grid size.
- A 111x111km variant existing at all suggests there is a real accuracy-per-compute curve to sit on, not a single fixed operating point.
- Because the model was co-trained on IBTrACS, its cyclone skill is tied to a curated observational record; treat it as a cyclone model with global weather capability, not a general model that happens to see storms.
Ensemble size is the product
WeatherNext uses Functional Generative Networks to produce ensembles of different predictions, which is how it represents forecast uncertainty. This is not described as a diffusion sampler in the source, and the distinction matters if you are budgeting for iterative denoising steps per member — the blog does not describe the sampling loop. What it does describe is scale: last year the system produced 50 predictions at a time, matching global physics models; this year it runs 1,000 members.
That jump exists for a specific reason. Fifty members give you a central estimate and a rough spread. One thousand members let you resolve rare but consequential scenarios — the post names rapid intensification, as seen with Melissa in 2025. A 1,000-member run over global atmospheric conditions produces localised probability maps of tropical storm to hurricane-force winds out to 15 days. If your downstream product is a threshold decision — evacuate, de-energise, reposition assets — the tail is the entire point and 50 members will not give it to you.
On cost, the one figure given is that a single 15-day forecast takes less than a minute on a TPU. The post does not name the TPU generation, does not state whether members batch efficiently, and does not give a wall-clock or dollar figure for a full 1,000-member run. Naively multiplying one minute by 1,000 is the wrong instinct and also the only arithmetic the blog supports, which is precisely why you should measure it yourself before committing to an operational cadence.
- Establish per-member marginal cost at your batch size before choosing an ensemble size; the 50-to-1,000 decision is a cost decision disguised as an accuracy decision.
- Decide whether you need 1,000 members always or only when a storm is active — cyclone season is bursty and the compute profile should be too.
- Storage and serving of 1,000 members x 15 days of global fields is a non-trivial pipeline in its own right; the post does not discuss output size.
Validating the skill claim
The comparison points DeepMind names are concrete, which makes them checkable. The accuracy plots show 3-day ECMWF-ENS track forecasts and 3-day HWRF intensity forecasts over time, with WeatherNext Cyclones placed as a step change on both. The evaluation window is historical cyclones from 2023 to 2024. Anything you build on top of this should reproduce against those same baselines and that same window before you extend the claim to your own use case.
- Separate the three claims — track, intensity, wind structure — and check each independently; a single model doing all three is the novel part and each may hold to a different degree.
- Confirm whether the ECMWF-ENS and HWRF baselines were run at their operational configurations for the 2023-2024 period, and at what ensemble size, before treating the gap as apples-to-apples.
- Distinguish the deterministic result from the probabilistic one; a 1,000-member system should be judged on calibration of the tails, not only on mean error.
- Treat the Hurricane Melissa case as an anecdote of operational usefulness, not as an evaluation result — it is one storm from the 2025 season, outside the reported 2023-2024 benchmark.
What is still open
The mechanism is unresolved by DeepMind's own account: nobody has a full explanation for why a 28x28km model matches or beats high-resolution specialised systems on intensity, and the team says so directly. That is an unusual and useful admission, and it means the failure modes are not yet mapped. A model whose skill source you cannot name is a model whose degradation you cannot anticipate.
The practical gaps to watch are licence terms and weight availability for both WeatherNext 2 and WeatherNext Cyclones, the sampling cost curve as ensemble size grows past 1,000, and whether the reported 24-hour lead-time gain survives independent evaluation on seasons outside 2023-2024. Until those land, the defensible position is that coarse-input generative ensembles are now a credible substitute for high-resolution physics on cyclone forecasting, and that the ensemble size, not the grid, is where your compute budget goes.
These daily notes are drafted by a model I run and operate myself — the same kind of pipeline this site is about — from sources published in the previous 24 hours, and every one lists what it read. The longer essays, the talks and the preprint are mine, written by hand.