L1 Penalty

The L1 Penalty is a regularisation term that adds the sum of the absolute values of a model's weights, scaled by a hyperparameter $\lambda$, to the loss: $\lambda \sum_i |w_i|$.

It pushes many weights to exactly zero, so it tends to produce sparse models. It's the penalty used in Lasso regression.

It's not the same as the L1 loss, which measures the absolute error of predictions. See Mean Absolute Error.