Multi-model ensembles (MMEs) are used to improve the forecasts of thermospheric neutral densities. A variety of algorithms for constructing the model weights for the MMEs are described and have been implemented including: performance weighting, independence weighting and non-negative least squares. Using both empirical and physics-based models, compared against in-situ CHAMP observations, the skill of each MME weighting approach has been tested in both solar minimum and maximum conditions. In both cases the MME performs better than any individual model. A non-negative least squares weighting for the MME on a set of bias corrected models provides a 68% and 50% reduction in the mean square error compared to the best model (Jacchia-Bowman 2008) in the solar minimum and maximum cases respectively.