WeatherNext 3 is Google's most advanced global weather forecasting model,
developed by [Google DeepMind](https://deepmind.google/) and Google Research.
By ingesting live geostationary satellite observations as a direct model input,
WeatherNext 3 is initialized every hour, producing forecasts at up to 0.05°
(5 km) spatial resolution with hourly timesteps. It also trains directly on
real-world weather station data, producing output that better matches what is
measured on the ground. For a detailed technical overview, see the
[research paper](https://arxiv.org/abs/2609.03582),
the [research and benchmarks](https://developers.google.com/weathernext/guides/research.md.txt) page, and the
[announcement blog post](https://blog.google/innovation-and-ai/models-and-research/google-deepmind/introducing-weathernext-3/).


[Get Forecast Data](https://developers.google.com/weathernext/guides/access-forecast.md.txt)
[Explore Variables](https://developers.google.com/weathernext/guides/models.md.txt#supported-variables)

## At a glance

| Attribute | Details |
| Release Date | August 2026 |
| Coverage | Global |
| Spatial Resolution | 0.05° (\~5 km) stations · 0.1° (\~10 km) gridded surface · 0.25° (\~25 km) pressure levels |
| Temporal Resolution | 1 hour (hourly timesteps) |
| Forecast Horizon | **15 days (360 hours)** for 6-hourly inits (00, 06, 12, 18 UTC) **48 hours** for interim inits (hourly from 2026; 03, 09, 15, 21 UTC only for 2024--2025) |
| Initialization Frequency | **2026 to present:** Every hour (24 inits per day) **2024--2025:** 8 inits per day (00, 03, 06, 09, 12, 15, 18, 21 UTC) |
| Ensemble Members | 64 |
| Architecture | Functional Generative Network (FGN) mesh transformer ([paper](https://arxiv.org/abs/2609.03582)) |
| Inputs | Live geostationary satellite mosaics ([SatWeave](https://developers.google.com/weathernext/guides/satweave.md.txt)) + ECMWF HRES analysis |
| Training Data | ERA5 / HRES-fc0, IMERG, [PARDIG](https://developers.google.com/weathernext/guides/pardig.md.txt), station observations, geostationary satellite mosaics ([SatWeave](https://developers.google.com/weathernext/guides/satweave.md.txt)) |
| Historical Forecasts | 2024 to present, for backtesting and evaluation (see [archive layout](https://developers.google.com/weathernext/guides/gcs.md.txt#directory-structure)) |
|---|---|

## Key features

- **Real-time satellite input:** WeatherNext 3 ingests a live global
  geostationary satellite mosaic ([SatWeave](https://developers.google.com/weathernext/guides/satweave.md.txt))
  as a direct model input, allowing it to be initialized **every hour**. This
  hourly refresh provides additional lead time when sudden weather fronts,
  storms, or precipitation systems develop rapidly.

- **Multi-resolution output:** Forecasts are generated at multiple spatial
  resolutions from a single forward pass:

  - **0.05° (\~5 km):** Station-trained 2m temperature and dew point.
  - **0.1° (\~10 km):** Gridded surface wind (10m and 100m), pressure, sea surface temperature, cloud layers, solar radiation fields, 1-hour IMERG precipitation, and 1-hour experimental satellite-radar precipitation.
  - **0.25° (\~25 km):** 3D atmospheric pressure levels (available in 6-hourly cycles: 00, 06, 12, 18 UTC).
- **Ground-truth station training:** Instead of training exclusively on
  reanalysis grids, WeatherNext 3 trains dedicated observational
  heads on raw weather station observations, producing 0.05° output that
  closely matches what is measured on the ground.

- **Precipitation:** WeatherNext 3 trains against three distinct
  precipitation data sources: ECMWF reanalysis, NASA's satellite-based IMERG,
  and Google's own satellite-radar precipitation reanalysis
  ([PARDIG](https://developers.google.com/weathernext/guides/pardig.md.txt)). This achieves up to a
  **50% reduction in Brier score and CRPS** compared to numerical weather
  prediction baselines when evaluated against IMERG observations.

- **Clean energy variables:** Outputs comprehensive parameters for
  renewable energy applications: 100-metre wind speeds for turbine-height
  forecasting, full cloud layer distributions, and complete solar irradiance
  components (SSRD, FDIR).

*** ** * ** ***

## Forecast cycles and horizons

WeatherNext 3 operates across distinct synoptic and interim initialization
cycles with varying forecast horizons and available variable scopes. Explore
initialization cycles and calculate coverage windows below:
<iframe src="https://developers.devsite.google/frame/weathernext/guides/models_7edcf41b3a41901385a507c93ba6df9ee72f3adf7f69654a328139606890fded.frame" class="framebox inherit-locale " allow="clipboard-write https://developers.devsite.google" allowfullscreen is-upgraded title="WeatherNext 3 Initialization and Horizon Visualizer"></iframe>

*** ** * ** ***

## Supported variables

WeatherNext 3 produces operational gridded surface forecasts, high-resolution
station observational head outputs, and 3D upper-air atmospheric fields.

### 1. Gridded surface variables (0.1° resolution)

1-hour step variables are available across all initialization runs (6-hourly and
interim hourly) on BigQuery, Earth Engine, and Google Cloud Storage.

| Variable Name | Description | Units | Resolution |
|---|---|---|---|
| Temperature \& Humidity ||||
| `temperature_2m` | 2 metre air temperature | K | 0.1° (\~10 km) |
| `dewpoint_temperature_2m` | 2 metre dew point temperature | K | 0.1° (\~10 km) |
| Wind (Surface \& Turbine Hub Height) ||||
| `wind_speed_10m` | 10 metre scalar wind speed (derived: √(u^2^ + v^2^)) | m/s | 0.1° (\~10 km) |
| `wind_speed_100m` | 100 metre scalar wind speed (derived: √(u^2^ + v^2^)) | m/s | 0.1° (\~10 km) |
| `u_component_of_wind_10m` | 10 metre U (eastward) wind component | m/s | 0.1° (\~10 km) |
| `v_component_of_wind_10m` | 10 metre V (northward) wind component | m/s | 0.1° (\~10 km) |
| `u_component_of_wind_100m` | 100 metre U (eastward) wind component | m/s | 0.1° (\~10 km) |
| `v_component_of_wind_100m` | 100 metre V (northward) wind component | m/s | 0.1° (\~10 km) |
| Solar Irradiance (Renewable Energy) ||||
| `surface_solar_radiation_downwards_1hr` | 1-hour surface downward solar radiation (SSRD / GHI) | J/m^2^ | 0.1° (\~10 km) |
| `total_sky_direct_solar_radiation_at_surface_1hr` | 1-hour total-sky direct solar radiation (FDIR / Direct Beam) | J/m^2^ | 0.1° (\~10 km) |
| Precipitation ||||
| `total_precipitation_1hr` | 1-hour accumulated precipitation (model native) | m | 0.1° (\~10 km) |
| `imerg_tp_1hr` | IMERG satellite-calibrated 1-hour precipitation | m | 0.1° (\~10 km) |
| `experimental_tp_1hr` | Experimental satellite-radar 1-hour precipitation ([PARDIG](https://developers.google.com/weathernext/guides/pardig.md.txt)-trained) | m | 0.1° (\~10 km) |
| Cloud Layers \& Cover ||||
| `total_cloud_cover` | Total atmospheric column cloud cover fraction | 0--1 | 0.1° (\~10 km) |
| `high_cloud_cover` | High cloud cover fraction | 0--1 | 0.1° (\~10 km) |
| `medium_cloud_cover` | Medium cloud cover fraction | 0--1 | 0.1° (\~10 km) |
| `low_cloud_cover` | Low cloud cover fraction | 0--1 | 0.1° (\~10 km) |
| Pressure \& Ocean ||||
| `mean_sea_level_pressure` | Mean sea level pressure | Pa | 0.1° (\~10 km) |
| `sea_surface_temperature` | Sea surface temperature | K | 0.1° (\~10 km) |

> [!NOTE]
> **Precipitation units:** All 1-hour precipitation variables (`total_precipitation_1hr`, `imerg_tp_1hr`, and `experimental_tp_1hr`) report 1-hour accumulated precipitation in **meters (m)** across all platforms. Multiply values by 1,000 to convert to millimetres (mm).

### 2. Station point variables (0.05° resolution)

Trained directly against surface weather station measurements. Available across
all inits on BigQuery, Earth Engine, and Google Cloud Storage.

| Variable Name | Description | Units | Resolution |
|---|---|---|---|
| `station_head_temperature_2m` | 2 metre air temperature at station locations | K | 0.05° (\~5 km) |
| `station_head_dewpoint_temperature_2m` | 2 metre dew point temperature at station locations | K | 0.05° (\~5 km) |

### 3. Atmospheric variables (0.25° resolution, pressure levels)

Available across 13 vertical pressure levels (**50, 100, 150, 200, 250, 300,
400, 500, 600, 700, 850, 925, 1000 hPa**) for 6-hourly inits (00, 06, 12, 18
UTC) exclusively on Google Cloud Storage (Full Ensemble Zarr).

| Variable Pattern | Description | Units | Resolution |
|---|---|---|---|
| `geopotential_{level}` | Geopotential (Φ; divide by 9.80665 m/s^2^ for geopotential height in meters) | m^2^/s^2^ | 0.25° (\~25 km) |
| `specific_humidity_{level}` | Specific humidity | kg/kg | 0.25° (\~25 km) |
| `temperature_{level}` | Air temperature | K | 0.25° (\~25 km) |
| `u_component_of_wind_{level}` | U (eastward) wind component | m/s | 0.25° (\~25 km) |
| `v_component_of_wind_{level}` | V (northward) wind component | m/s | 0.25° (\~25 km) |
| `vertical_velocity_{level}` | Vertical velocity (ω) | Pa/s | 0.25° (\~25 km) |

Replace `{level}` with the pressure level in hPa (for example,
`geopotential_500` represents geopotential at 500 hPa). In unflattened raw Zarr
stores, variables are indexed along the `level` coordinate dimension.

*** ** * ** ***

## How to access WeatherNext 3 data

WeatherNext 3 forecast data is available through BigQuery, Earth Engine, and
Google Cloud Storage. Visit the
[quick start guide](https://developers.google.com/weathernext/guides/access-forecast.md.txt) to compare access
options, request allowlist access, and run starter queries.

> [!NOTE]
> WeatherNext 3 predictions are provided "as-is" for informational purposes
> and are not official severe weather warnings. Always defer to official alerts
> from emergency authorities. For more details, see the
> [Terms of Service and
> Disclaimers](https://developers.google.com/weathernext/guides/disclaimers).

*** ** * ** ***

## Migrating from WeatherNext 2

> [!NOTE]
> **Variable naming convention change:** WeatherNext 3 adopts a new variable naming convention where the height indicator is placed as a suffix instead of a prefix. If you are migrating existing queries or pipelines from WeatherNext 2, update your variable references accordingly.

Key variable name changes:

| WeatherNext 2 | WeatherNext 3 |
|---|---|
| `2m_temperature` | `temperature_2m` |
| `2m_dewpoint_temperature` | `dewpoint_temperature_2m` |
| `10m_u_component_of_wind` | `u_component_of_wind_10m` |
| `10m_v_component_of_wind` | `v_component_of_wind_10m` |
| `100m_u_component_of_wind` | `u_component_of_wind_100m` |
| `100m_v_component_of_wind` | `v_component_of_wind_100m` |

Variables without a height indicator (for example, `mean_sea_level_pressure`,
`total_cloud_cover`, `geopotential`) retain the same names across both versions.

> [!WARNING]
> **Precipitation accumulation interval change:** WeatherNext 2 reported 6-hour accumulated precipitation (`total_precipitation`). WeatherNext 3 reports **1-hour accumulated precipitation** (`total_precipitation_1hr`, `imerg_tp_1hr`, and `experimental_tp_1hr` all in meters). If your data pipelines aggregate precipitation over daily or multi-day periods, adjust your integration math to sum 1-hour steps rather than 6-hour steps.

*** ** * ** ***

## Model lineage \& earlier versions

Google has invested in machine learning weather research for many years,
producing a series of models that have consistently pushed the state of the art:

- **WeatherNext 3 (August 2026):** Flagship operational model. FGN mesh transformer with live satellite assimilation, multi-resolution output (0.05° stations, 0.1° surface, 0.25° pressure levels), and hourly initialization. Recommended for all new projects.
- **[WeatherNext 2](https://developers.google.com/weathernext/guides/models-wn2.md.txt)** (June 2025): Ensemble model based on the Functional Generative Network (FGN) architecture. 0.25° resolution, 6-hourly timesteps, 64 ensemble members.
- **[WeatherNext 1 Gen](https://developers.google.com/weathernext/guides/models-wn-gen.md.txt)** (Dec 2024): Probabilistic diffusion-based ensemble model (GenCast). 0.25° resolution, 12-hourly timesteps, 50 ensemble members.
- **[WeatherNext 1 Graph](https://developers.google.com/weathernext/guides/models-wn-graph.md.txt)** (Nov 2023): Deterministic Graph Neural Network model (GraphCast). 0.25° resolution, 6-hourly timesteps.

Older models remain available for research and benchmarking.

*** ** * ** ***

## Terms of service and disclaimers

WeatherNext 3 is an experimental AI forecasting system under active
research. Predictions and model outputs are provided "as-is" for informational
and research purposes and do not constitute official weather forecasts, watches,
or warnings. Users are responsible for their use of model outputs and data.
For the protection of life and property, never rely on WeatherNext 3 as a sole
source of information and always defer to official alerts and advisories from
your national meteorological service and local emergency authorities.

### Data licensing and terms of use

- **Real-Time Data \& Forecasts:** Any forecasts, model predictions, or data relating to a time less than 1 hour ago and the future are governed by the [GDM Real-Time Weather Forecasting Experimental Data Terms of
  Use](https://storage.googleapis.com/weathernext-public/terms-of-use.pdf).
- **Historical Data:** Any data relating to a time 1 hour ago or more is licensed under the [Creative Commons Attribution International License, Version 4.0 (CC BY 4.0)](https://creativecommons.org/licenses/by/4.0/). When real-time data ages past the 1-hour threshold, it automatically transitions to CC BY 4.0.

For complete terms, liability provisions, and acceptable use restrictions, see
the [Terms of Service and Disclaimers](https://developers.google.com/weathernext/guides/disclaimers.md.txt) page.

*** ** * ** ***

See the [Glossary](https://developers.google.com/weathernext/guides/glossary.md.txt) for definitions of
meteorological terms, data formats, and Google Cloud platforms.