I too am curious to hear the results you see when testing the classification model directly in Create ML app (assuming that's where you created the model). Also, I'm curious to know how much training data you have and what string values you're testing with. Aside from that, there is a performance improvement you can make here. You are loading the MLModel on every prediction which takes time. I'd load the model once (on first prediction) and reuse it with any subsequent prediction.