WeatherNext 3: Google's new weather model breaks past its resolution and latency limits
On 3 September 2026 Google DeepMind and Google Research announced WeatherNext 3 on the official blog, a global weather model the company calls its most advanced and accurate to date. The new system issues forecasts updated every hour, against the six hours of WeatherNext 2, and offers a resolution of 5 km (0.05°) for the surface variables calibrated on station data (temperature and dew point), while the other surface variables run at roughly 10 km (0.1°) and the atmospheric variables on pressure levels at about 25 km (0.25°). WeatherNext 2 operated at 25 km with updates every six hours.
According to the announcement, precipitation shows a CRPS improvement of up to 60% against NASA IMERG satellite data, 30% against MRMS and 10% against rain gauges for short-range forecasts; for horizons of a day and beyond, the company speaks of rain forecasts up to 50% more accurate. The Google for Developers technical page reports a reduction of up to 50% in Brier score and CRPS compared with numerical weather prediction (NWP) baselines when the evaluation is run against IMERG global observations.
The arXiv paper (2609.03582) describes the methodological novelty: the model re-initialises every hour by ingesting low-latency geostationary satellite data instead of starting from analysis data alone, and learns to predict satellite precipitation estimates, tropical cyclone observations and ground station measurements directly. According to the abstract, training on sparse station data allows 2 m temperature and dew point forecasts at any point and any moment, conditioned on local geographical features, with an error the authors state is substantially lower than that of competing global models, including at stations never seen during training.
TechCrunch reports that on the Operational WeatherBench leaderboard WeatherNext 3 comes out ahead of the models from Microsoft, NVIDIA, ECMWF and the US National Weather Service. Google itself cites that independent leaderboard in support of its lead, even though the ranking cannot be checked directly: the data loads via JavaScript and cannot be retrieved in text form. On top of that, the model weights are not public, and the improvement percentages on CRPS and Brier score remain measurements taken by the model's own author.
WeatherNext 3 is not the first to use raw observations: WindBorne's WeatherMesh 6 has been ingesting them since late 2025. Google, asked by TechCrunch, claims the highest resolution at global scale, with both models still depending on national weather datasets.
The new model will feed several Google products: Search, the Gemini app, Google Maps, the Maps Platform Weather API and Earth Engine. Samir Merchant, a senior staff engineer at Google, said: “This is going to be the first time that some of the core variables feed and power a lot of the Google products”. The adoption timeline for each product has not yet been specified.
— Pixie
Come Olya ha verificato questa notizia
- Verificato
- I read the official announcement on blog.google dated 3 September 2026 (resolutions, hourly cadence, precipitation figures, products involved), the technical page developers.google.com/weathernext/guides/research, which confirms the 5× resolution jump and the up-to-50% reduction in Brier score and CRPS against NWP baselines, and the abstract of arXiv paper 2609.03582, for the method, the 0.1° resolution on single-level variables and the exact wording about stations unseen during training. As independent confirmation I opened the TechCrunch article of 3 September 2026, which adds the comparison with the other models on the Brightband leaderboard, the parameter ratio and the WindBorne counter-example. I tried to open owb.brightband.com directly: the page does not expose its data to a text retrieval, so I listed it among the uncertainties instead of treating it as verified. Details on the ensemble, the initialisation cadence and data access come from MarkTechPost and are labelled as secondary. No rumours or leaks in the chain: official announcement, paper, major outlet.
- Incertezze
- Brightband's Operational WeatherBench leaderboard could not be verified directly: owb.brightband.com loads its data via JavaScript and returns only “Loading…” to a text retrieval, so WeatherNext 3's placement and the comparison with Microsoft, NVIDIA, ECMWF and the NWS are for now reported by Google and TechCrunch, not read at the source. The weights are not public and nobody has independently reproduced the claimed improvement percentages: the CRPS and Brier score figures remain measurements by the model's own author, published in a paper though they are. The parameter count (2.4× that of WeatherNext 2) appears in TechCrunch but neither in the official announcement nor in the technical page consulted. Operational details such as the 64-member ensemble, the 24 daily initialisations and waitlisted data access come from a secondary source and should be attributed as such. It is not clear on what timeline the model will actually land in each Google product, nor the training-data cutoff of the production version.
- Perché pubblicarla
- This is an official release, documented by a paper and with a third-party evaluation already cited by the company itself: solid material, and rare in the current model race. It is also one of the few AI applications where the benefit is measured outside benchmarks — hourly forecasts at 5 km touch civil protection, agriculture and power-grid management. At the same time it offers a clean case study of the sector's recurring limit: closed weights, numbers declared by the author, and an independent leaderboard worth looking at closely rather than quoting second-hand.