WeatherNext models are global medium-range forecasting models with broad
applications across industries. If your work relies on models such as
ECMWF HRES/ENS/AIFS or NOAA GFS, WeatherNext offers a compelling alternative.

## Benefits

- **Accuracy:** WeatherNext 3 outperforms leading operational NWP systems
  across the majority of variables, lead times, and pressure levels. See the
  [research and benchmarks](https://developers.google.com/weathernext/guides/research.md.txt) page for detailed
  benchmarks.

- **Probabilistic forecasting:** WeatherNext 3 generates 64-member ensembles
  to characterize predictive uncertainty and estimate risk. (For custom
  ensemble sizing on demand,
  [WeatherNext 2 managed inference](https://developers.google.com/weathernext/guides/access-vmg.md.txt)
  is available on Vertex AI Model Garden.)

- **Speed:** A full 15-day global forecast can be generated in minutes on a
  single TPU, enabling rapid iteration and near real-time applications. For
  forecast availability details, see the [dissemination
  schedule](https://developers.google.com/weathernext/guides/dissemination).

- **High resolution:** WeatherNext 3 produces forecasts at up to 0.05° (\~5 km)
  spatial resolution for station-calibrated surface variables
  (`station_head_temperature_2m` and `station_head_dewpoint_temperature_2m`),
  and 0.1° (\~10 km) for core gridded surface fields (wind, pressure,
  precipitation, clouds, solar) with hourly timesteps. This provides a
  2.5× to 5× resolution improvement over WeatherNext 2's 0.25°
  grid, resolving local features like valleys, coastlines, and topography.

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

- **Clean energy parameters:** WeatherNext 3 directly outputs 100-metre wind
  speeds, cloud cover, and solar radiation components (SSRD, FDIR) for
  renewable energy applications.

- **Downstream integration:** WeatherNext outputs can serve as inputs to other
  systems and models, such as ensemble post-processing pipelines, [flood
  forecasting](https://research.google/blog/a-flood-forecasting-ai-model-trained-and-evaluated-globally/)
  models, and specialized regional downscaling models.

## Limitations

- **Reanalysis targeting:** WeatherNext models are trained against global
  reanalysis data (ERA5 / HRES-fc0), which has limited resolution and carries
  its own biases. Reanalysis does not always match what you would measure on
  the ground, particularly for localized variables. For applications requiring
  close agreement with surface observations, bias-correction techniques may be
  needed. WeatherNext 3 partially addresses this by also training on raw
  weather station observations for key surface variables.

- **Blurring:** Deterministic ML models like WeatherNext Graph tend to produce
  forecasts that become progressively smoother at longer lead times. This
  happens because the model learns to predict the average of possible
  outcomes, which washes out fine-scale detail as uncertainty grows.
  Probabilistic ensemble models (WeatherNext 2 and 3) are less affected
  because individual ensemble members preserve sharper structures.

- **Precipitation data quality:** Model precipitation outputs are influenced
  by the quality and biases of their training data. ERA5 precipitation has
  known limitations and is often excluded from headline evaluations (see
  [research papers](https://developers.google.com/weathernext/guides/research.md.txt) for details). WeatherNext 3
  mitigates this by training against multiple precipitation sources including
  satellite-based observations.

- **Artifacts in predictions:** As demonstrated in [Alet et al.
  (2025)](https://arxiv.org/abs/2506.10772), despite being trained to
  optimize marginal skill, WeatherNext models produce skillful predictions of
  joint structures. Nevertheless, some artifacts are visible in individual
  output samples, reflecting the underlying mesh structure of the model. These
  artifacts were noted already for WeatherNext 2 and are still present in
  WeatherNext 3. While visible to some degree in most variables, they are more
  severe in some, in particular the station output and precipitation.

  In the experimental satellite-radar ([PARDIG](https://developers.google.com/weathernext/guides/pardig.md.txt))
  head output, individual samples show clear hexagonal patterns. The ensemble
  mean still shows these while they are less noticeable in the median. The
  experimental satellite-radar head shows these artifacts more strongly than
  the IMERG head. This is due to its distribution being sharper, which leads
  to overall better performance but also more evident artifacts in the
  individual members, compared to IMERG's smoother distribution (see
  [research and benchmarks](https://developers.google.com/weathernext/guides/research.md.txt)). Despite these
  unrealistic patterns, WeatherNext 3 still performs strongly on pooled
  metrics that take into account spatial correlations. This is important to
  make reliable probabilistic predictions of cumulative precipitation over
  larger areas, used for example in flood forecasting.

  WeatherNext 3 produces outputs at time steps smaller than the outer 6-hour
  time step. For variables that are not auto-regressively predicted, such as
  the station head, distinct jumps can occur across the 6-hour windows. In
  addition, the station head also exhibits a per-sample bias, where one sample
  will tend to be warmer or colder globally compared to the ensemble mean. The
  global bias changes with each 6-hour time step as a new noise vector is
  drawn. This behaviour shows another way the model produces realistic
  marginal distributions while approximating the covariance structure.
  However, when looking at marginal statistics such as quantiles, the
  discontinuities are relatively minor. Since many applications rely mostly on
  ensemble statistics (e.g., the ensemble mean or exceedance probabilities),
  the forecasts are nevertheless widely usable.