Unfortunately not much. Can't you run even quantized models?