A convection-allowing model (CAM) is a computer weather model with grid points close enough together, typically 4 km (about 2.5 miles) apart or closer, that thunderstorms can form directly in the simulation. Coarser models have to estimate what storms would do with a simplified scheme, while a CAM builds its own storms from the model's equations.
Resolving storms instead of estimating them
A model cannot depict features that are small compared with the spacing between its grid points. When that spacing is wide, a thunderstorm is too small to simulate, so the model uses a convective parameterization, a shortcut that estimates the rain and heating storms would produce.
The American Meteorological Society describes a CAM as having enough resolution to handle the main processes inside deep convective storms explicitly. These models are also called convection-permitting models.
What they add for forecasters
Because individual storms appear in the output, a CAM can show simulated radar reflectivity. A forecaster can see whether the model favors scattered cells, a line or a large cluster, and roughly when storms start.
Researchers at the Storm Prediction Center (SPC) and the National Severe Storms Laboratory (NSSL) found that CAMs add value with explicit predictions of extreme convective features where older models only predict convective rainfall. Hourly maximum fields, such as updraft helicity for rotating storms, record the strongest value between output times so short-lived peaks are not missed.
Examples
NOAA's High-Resolution Rapid Refresh (HRRR) is a 3-km, hourly updated CAM. The High Resolution Ensemble Forecast (HREF) combines several models running at about 3 km.
As of October 2026, the National Weather Service plans to retire HREF and several other regional systems on 10/14/26, replacing them with the Rapid Refresh Forecast System (RRFS) and its ensemble.
Where they fall short
The same SPC and NSSL work notes that CAMs often miss the timing, strength and sometimes the existence of individual storms. They can also place the most intense activity in the wrong part of a broader area of storms.
A 3-km grid is still far too coarse to depict a tornado. Compare successive runs and ensembles before trusting any single simulated storm.