Such models
have the very minimal number of layers (e.g., 2), vocab size (e.g., 1000), etc.