Google DeepMind and Google Research announced their global weather forecasting model "WeatherNext 3" on September 3, 2026. The model feeds in geostationary satellite observations directly and generates new forecasts every hour. Some outputs, such as temperature, are now rendered on a finer grid of roughly 5km, and the model began rolling out to Google Search, the Gemini app, and Google Maps the same day.
This is an update that brings new observations into forecasts faster, moving predictions closer to what is actually measured on the ground. However, the headline features—"hourly," "5km," and "precipitation forecasts improved by up to 50%"—each apply to different targets and conditions. Cross-referencing the published research report with the operational specifications reveals which forecasts have actually become finer-grained, and what they can and cannot be used to judge.
What changes with direct satellite input
WeatherNext 3 uses satellite observations as input in addition to the "analysis" data derived from numerical weather prediction—the estimated state of the global atmosphere produced by combining observations from around the world with a numerical prediction model. Previous global AI models improved their performance by training on this neatly gridded analysis data and using it to predict future states.
The problem is that waiting for analysis data to be finalized means the information fed into the forecast is already somewhat stale. The model also inherits any biases present in the surface temperature and precipitation values within that analysis data. WeatherNext 3 adds satellite observations available with roughly a one-hour delay, incorporating the latest changes in clouds and water vapor. It is designed to continue using numerical weather prediction analysis data while updating forecasts with newer observations.
Internally, it uses the Functional Generative Network (FGN) inherited from the previous generation. By feeding the model different noise inputs, it generates multiple possible futures and represents uncertainty as an ensemble of 64 forecasts. Because inputs at different resolutions are gathered into a shared internal representation and outputs are split by weather variable, the same model can handle both broad atmospheric predictions and fine-grained surface forecasts.
The changes on the surface side are substantial too. The output component for weather stations learns directly from the temperature and dew point readings measured at actual observation stations. It also uses information such as elevation to reflect temperature variation along coastlines and in mountainous terrain. The research team held out 5% of stations from training to verify performance at unseen locations. That said, biases arise in regions with sparse observations, which are supplemented using pseudo-observation data generated from analysis data. The ability to produce output globally is not the same as being able to verify it with equal density everywhere.
Which weather variables actually get the 5km forecast?
The resolution of WeatherNext 3's operational data varies by weather variable. The 5km output applies to temperature and dew point, which are trained on weather-station data; precipitation and near-surface wind are delivered on a roughly 10km grid, while upper-air variables use a roughly 25km grid.
| Output type | Grid resolution | Main target variables | Forecast cycle |
|---|---|---|---|
| Surface predictions trained on station data | 0.05°, ~5km | 2m air temperature and dew point | Hourly |
| General surface predictions | 0.1°, ~10km | Precipitation, near-surface wind, cloud cover, solar radiation, etc. | Hourly |
| Upper-air predictions | 0.25°, ~25km | Temperature, wind, etc. at each pressure level | Main 6-hour cycle |
Dew point is a measure of how humid the air is. Compared to the roughly 25km grid that underpins the previous-generation WeatherNext 2, the grid spacing for the weather-station output is about one-fifth as wide. This is a comparison of the spacing between forecast points—it does not mean the error itself has shrunk to one-fifth. Nor has rain forecasting uniformly become 5km-resolution.
There's also a distinction in forecast range. The main cycles, which start at 00:00, 06:00, 12:00, and 18:00 UTC, cover up to 15 days ahead, while the hourly cycles in between predict surface and station-level variables up to 48 hours ahead. The previous generation converted its basic 6-hour forecasts into 1-hour intervals using a separate model, but WeatherNext 3 directly generates output at 1-hour intervals. Both the update frequency and the time granularity within the forecast have changed.
Moreover, hourly updates do not mean instant delivery. According to the official delivery schedule, a forecast nominally initialized at 00:00 UTC arrives in Cloud Storage around 07:45, while BigQuery and Earth Engine deliver it around 08:10. The nominal initialization time, the most recent satellite observation included in the input, and the actual delivery time are all different. Anyone using this operationally needs to compare forecasts based on how many hours ahead they are relative to when the data actually arrived.
How to read the "up to 50% improvement" in precipitation
Google states that precipitation forecasts more than a day out are up to 50% more accurate. The developer-facing evaluation notes explain that, when measured against the satellite-derived precipitation estimate dataset "IMERG," the model reduced CRPS and Brier score by up to 50% relative to a numerical weather prediction baseline. The figure in the product description and the individual evaluation results should not be treated as the same measurement without qualification.
CRPS measures the discrepancy between a predicted probability distribution and the actual observation—lower is better. Brier score likewise measures the error in probabilistic forecasts, such as "the probability that rainfall will exceed a certain amount." It does not mean "the probability of rain being correct rises by 50 percentage points."
The research report by Stephan Rasp and colleagues breaks the evaluation down further. The main verification used a model trained on data through the end of 2023 and was conducted across the full year of 2024. For precipitation, WeatherNext 2 and the European Centre for Medium-Range Weather Forecasts' (ECMWF) ENS were used as comparison baselines, and the largest CRPS reductions were reported at short forecast lead times.
| Reference observation dataset | Precipitation accumulation window | Reported maximum reduction | Conditions to note |
|---|---|---|---|
| Global satellite estimate IMERG | 6 hours | 60% | Also used as a training target, so not a fully independent verification dataset |
| MRMS (mainly U.S. ground radar, etc.) | 6 hours | 30% | Limited to one region, differing from the global evaluation |
| Rain gauges | 24 hours | 10% | Station distribution is uneven, and the accumulation window differs from the other rows |
Each row reflects a maximum value under different conditions—they are not a single, uniform improvement rate against the same forecast time and the same comparison target. A large improvement measured against a satellite estimate does not automatically translate into the same percentage improvement measured against ground rain gauges. On the other hand, the fact that improvements were also reported against MRMS and rain gauges suggests value that cannot be explained solely by the model fitting closely to data resembling its own training targets.
Generalizing to heavy rainfall requires even more caution. The Brier score evaluation broken down by rainfall intensity was limited to up to 4mm over 6 hours, and extreme precipitation events—being rare—were flagged as a topic for future work. Improvements in everyday "will it rain or not" forecasting cannot be read as a guarantee of predictive performance for extreme downpours or flooding.
There is also a separate evaluation using the operational version. A model trained on data through June 30, 2026 was compared against ECMWF's AI model "AIFS ENS v2" and others over a six-week period from July 1 to August 11. However, the research team itself cautions against over-interpreting small differences observed over such a short period. The research report is a PDF published by Google, and no peer-reviewed journal publication is noted.
The unnatural patterns that averaging hides
Individual WeatherNext 3 forecasts can sometimes show hexagonal patterns that reflect the internal computational grid. The research report notes that this is especially noticeable in precipitation and weather-station outputs. For station-level outputs and similar variables, unnatural jumps in time can also appear at the boundaries of the internal 6-hour segments.
Training that improves the probability distribution at each individual forecast point does not fully guarantee consistency with neighboring points or continuity across time. As a result, even when error metrics look good, individual generated weather maps for the future can still contain unnatural artifacts. Each generated forecast map needs to be checked for consistency across both space and time, not just point-by-point accuracy.
That said, the research reports good performance in precipitation evaluations aggregated over wide areas, and states that these artifacts are reduced when looking at medians or quantiles. A quantile represents where a value falls when predictions are sorted from smallest to largest. Depending on whether day-to-day decisions rely on averages, the probability of exceeding a certain rainfall threshold, or the time evolution of individual forecasts, the weaknesses that matter will differ.
For power generation planning, the retrieval method matters too
WeatherNext 3 forecasts wind and solar radiation at a height of 100m. For companies managing wind turbine operations or solar power generation, these are variables that are more directly useful than a surface weather icon. The addition of cloud cover forecasts gives more material to cross-reference against actual generation output at individual facilities.
How the data is retrieved differs by use case. Cloud Storage provides access to the full set of 64 ensemble members as well as upper-air data. BigQuery and Earth Engine offer pre-computed statistics, such as surface averages and quantiles. If you're linking forecasts to operational data for a specific site, starting with the statistical summaries is easier; if you need to examine situations where multiple sites experience bad weather simultaneously, or track changes over time, working with the full ensemble makes more sense.
Access to real-time operational data requires an application, with approval typically taking 5–7 business days. Historical data for times more than an hour in the past is licensed under CC BY 4.0, while real-time data covering the near past and future is governed by separate terms. WeatherNext 3 itself is not open source, distinguishing it from the previous-generation WeatherNext 2 and the cyclone model, whose code and weights were released publicly.
While the public announcement came on September 3, developer release notes list WeatherNext 3's release date as August 31. Rollout into Google products, data access, and the ability to run the model yourself are three separate forms of availability. For disaster preparedness, warnings should still be confirmed through official meteorological authorities.
To verify the value of hourly updates, one needs to measure, based on the actual time data arrives, how decision-making improves relative to existing forecasts for a specific region and season. For a power plant, the benefit of ingesting observations faster only becomes real once the wind and solar radiation forecasts for the times the facility actually needs are checked against real outcomes, and forecast uncertainty itself is incorporated into operational planning.