Many developing countries are adopting inflation targeting regimes to guide monetary policy decisions. In such countries the share of food in the consumption basket is high and policy makers often employ total inflation (as opposed to core inflation) to set inflationary targets. Therefore, central banks need to develop reliable models to forecast food inflation. Our literature review suggests that little has been done in the construction of models to forecast short-run food inflation in developing countries. We develop a model to improve short-run food inflation forecasts in Colombia. The model disaggregates food items according to economic theory and employs Flexible Least Squares given the presence of structural changes in the inflation series. We compare the performance of this new model to current models employed by the central bank. Next, we apply econometric methods to combine forecasts from alternative models and test whether such combination outperforms individual models. Our results indicate that forecasts can be improved by classifying food basket items according to unprocessed, processed and food away from home and by employing forecast combination techniques.