The intrinsic variability of the atmospheric system is historically reproduced by ensembles of forecasts based on numerical weather prediction. However, the computational cost of running such ensembles based on perturbed initial conditions is prohibitive. Recent advances in machine learning (ML)-based emulators for medium range weather forecasting have opened up new opportunities. While these methods require large amounts of training data and are computationally expensive during training, the inference/forecast step is computationally cheap. We propose a novel approach of perturbing pre-trained ML emulators. As it has been suggested that initial condition perturbations only work to a limited extent in ML emulators, we propose to perturb the latent spaces of these emulators directly, by adding noise to weight tensors. One advantage of this approach is that the perturbations can be applied iteratively. Thereby, the resulting probability distribution of the ensemble members can be adjusted to serve a specific need. First results suggest that introducing such perturbations allows the previously deterministic emulator to create a probabilistic ensemble weather forecast. These forecasts are thoroughly evaluated and compared against measurements from MeteoSwiss (the Swiss national weather service). The error growth and propagation of the perturbations are subject to careful analysis.