In a setting where multiple spectroscopic instruments are used for the same measurements it may be convenient to develop the calibration model on a single instrument and then transfer this model to the other instruments. In the ideal scenario, all instruments provide the same predictions for the same samples using the transferred model. However, sometimes the success of a model transfer is evaluated by comparing the transferred model predictions with the reference values. This is not optimal, as uncertainties in the reference method will impact the evaluation. This paper proposes a new method for calibration model transfer evaluation. The new method is based on comparing predictions from different instruments, rather than comparing predictions and reference values. A total of 75 flour samples were available for the study. All samples were measured on ten near infrared (NIR) instruments from two instrumental platforms, five NIR instruments from each platform. Protein content was quantified for all 75 samples and used as the reference variable during modelling by partial least squares regression. By adding artificial noise to first the spectroscopic measurements and then the reference values, this paper highlights the problems of including reference values in the evaluation of a model transfer, as uncertainties in the reference method impact the evaluation. At the same time, this paper highlights the power of the proposed model transfer evaluation, which is based on comparing predictions obtained from the different instruments. In this way, the impact of uncertainties originating from the reference method is minimised.