WeatherNext 3 had collected 227 Hacker News points when it reached our news brief, but the number developers should keep beside it is 48 hours. Google's new weather model can start a global forecast every hour, yet 20 of those 24 daily runs end after two days. Only the runs initialized at 00, 06, 12 and 18 UTC extend to the advertised 15-day horizon, according to the WeatherNext 3 developer documentation. That split matters when a pipeline has to combine fresh short-range updates with planning several days ahead.
Google describes WeatherNext 3 as the first global weather model to generate a forecast every hour. The company is putting its output into Search, Maps and Gemini, and it has made forecast data available through BigQuery, Earth Engine and Google Cloud Storage. The launch page also labels the system experimental and tells people to use their national weather service or local meteorological agency for official warnings.
The fresher input is the main change
Most AI weather systems begin with an analysis: a gridded estimate of the atmosphere assembled from observations by a numerical forecasting system. The WeatherNext 3 paper says global operational analyses are produced every six hours and can be six to 12 hours old by the time they are available. WeatherNext 3 still consumes ECMWF HRES analysis data, so it has not discarded that physical forecasting chain. It adds the latest geostationary satellite mosaic, whose operational delay is roughly one hour, then issues an updated forecast each hour. The research paper says those satellite inputs are at least five hours fresher than the newest analysis available to the model.
The difference is easiest to see in the authors' 14:00 UTC example. At that time, WeatherNext 3 can ingest satellite mosaics from 08:00 through 13:00, while the newest HRES analysis comes from 06:00 and contributes a forecast valid at 08:00. The satellite mosaic has 11 channels at 0.1-degree resolution across visible and infrared bands, carrying information about clouds, water vapour and temperature. This extra stream gives the model something recent to work with between the conventional analysis cycles, rather than pretending those cycles no longer exist.
WeatherNext 3 uses a Functional Generative Network built around an encode-process-decode mesh transformer. It samples 64 ensemble members by injecting noise, producing a distribution of possible weather paths instead of a single answer. The model works in six-hour outer steps, but it predicts many surface fields and precipitation at one-hour intervals within each step. The technical description says the four full daily cycles run to 360 hours; the 20 interim cycles cover the next 48 hours.
That design makes the hourly cadence useful for fast-developing rain and storm systems without multiplying every long forecast by six. In the paper's latency-adjusted precipitation test, hourly initialization gained about two to three hours of useful lead time over six-hour initialization. A developer polling for the newest run therefore needs to treat initialization_time and forecast horizon as part of the data contract, not assume that every new run replaces the whole 15-day series.
Five kilometres applies to two variables
Google's variable guide lists station-trained temperature and dew point at 0.05 degrees, or about 5 kilometres. Gridded surface fields such as wind, pressure, cloud cover, sea-surface temperature and hourly rain are 0.1 degrees, about 10 kilometres. Three-dimensional atmospheric fields across 13 pressure levels remain at 0.25 degrees, about 25 kilometres, and are available only in the four six-hourly cycles through the full-ensemble Zarr files in Cloud Storage.
For energy applications, the 0.1-degree surface product includes wind at 100 metres, cloud layers and solar irradiance. Those are directly useful inputs for estimating wind and solar production. They are forecasts at grid scale, though, and the documentation does not turn a 10-kilometre wind cell into a turbine-level prediction. Local terrain, equipment behaviour and site measurements still sit downstream of the global model.
The 5-kilometre station head is unusual because it learns from irregular station observations and combines its internal field with elevation and a land-or-sea indicator. During inference, the system can query temperature and dew point at arbitrary locations and times; Google publishes that head on a 0.05-degree hourly grid. The paper says querying exact station coordinates made only a marginal difference in its tests, which puts a useful limit on how much to read into the finer output spacing.
Existing WeatherNext pipelines need small but important edits
Google's public developer route is a data product. BigQuery and Earth Engine provide the hourly surface and station fields, while Cloud Storage carries the fuller ensemble files and pressure-level variables. Historical forecasts from 2026 are available for backtesting, and Google says it is backfilling 2024 and 2025. Live forecasts and data less than an hour old use Google's experimental weather-data terms; historical data becomes available under CC BY 4.0 once it passes that threshold.
WeatherNext 2 users cannot swap table names and walk away. Version 3 moves height indicators to the end of variable names and changes precipitation from six-hour totals to one-hour accumulations, still reported in metres. The migration looks small in code, but missed aggregation or unit conversion would produce a bad answer without throwing an error. Google's documented examples reduce to this pair of changes:
WeatherNext 2: 2m_temperature, total_precipitation # 6-hour rain
WeatherNext 3: temperature_2m, total_precipitation_1hr # 1-hour rain, metres
Daily rainfall now requires summing 24 hourly values and multiplying metres by 1,000 for millimetres. The developer guide also warns that six-hour accumulation variables are confined to the four synoptic cycles and the full Zarr data in Cloud Storage. A migration test should cover both an interim run and a full run so that a query does not fail only when its horizon or variable set changes.
Google tested one year, then six live weeks
The broadest evaluation used every day of 2024 and a research model trained only through the end of 2023. On upper-atmosphere variables, the authors report roughly a 5% improvement over WeatherNext 2, which they estimate as six more hours of lead time at the same skill. For short-range temperature at held-out METAR stations, the station head reduced continuous ranked probability score by as much as 30% against WeatherNext 2 and 40% against ECMWF's traditional ensemble. These are the Google team's tests, reported in a preprint submitted on September 3, rather than an independent operational audit.
Rain produced the largest percentages and the most reasons for caution. At early lead times, the model's PARDIG precipitation head cut the same error measure by up to 60% when judged against NASA IMERG, 30% against US radar data and 10% against rain gauges. The gains were largest near the rain-or-no-rain threshold. At heavier rain rates, the paper says the available Brier scores become small and less reliable at longer lead times, so the biggest percentage should not be read as a blanket 60% gain for severe rainfall.
Google also compared its production model with ECMWF's AIFS ENS v2 over six weeks, from July 1 to August 11, 2026. WeatherNext 3 averaged about 10% better on upper-level variables during the first forecast week. The authors explicitly warn against over-reading smaller differences from that short period. They also note that the systems were evaluated against slightly different analysis ground truths and that the lower-resolution AIFS output had to be interpolated for some comparisons.
Station coverage leaves another seam. Training combines about 5,000 METAR airport stations, a Mesonet collection that had roughly 15,000 stations in 2024, and ship and buoy reports. The Mesonet data is concentrated in Europe and North America. The team found biases in thinly observed areas including the Andes, Himalayas and some high-latitude oceans, then added 2,000 synthetic station locations per hour filled from analysis data to reduce those biases. Five percent of the real stations were held out consistently for the spatial-generalization test.
Individual ensemble samples have visible defects too. The paper shows hexagonal mesh patterns, especially in precipitation and station outputs, plus temperature jumps where one six-hour model step meets the next. Medians and probability thresholds suppress much of that structure, but applications that animate a single member or treat it as a coherent scenario could expose it. Cyclone predictions improved modestly on track and intensity, while their ensemble spread was narrower than WeatherNext 2's, a possible sign of overfitting according to the authors.
Google still needs to finish the 2024 and 2025 historical backfill. An operational evaluation longer than the current six weeks would carry more weight, especially if it reports whether the mesh artifacts and six-hour boundary jumps have improved. WeatherNext 3 currently makes the strongest case as a frequently refreshed probability feed for tested downstream systems. Its own documentation says it is not an official warning service, and no benchmark changes that boundary.