Presented at the 2013 SMOS-Aquarius Science WorkshopThe Soil Moisture and Ocean Salinity (SMOS) satellite has measured the L-band brightness temperature of the Earth over three years. The payload instrument MIRAS (Microwave Imaging Radiometer using Aperture Synthesis) measures two-dimensional brightness temperature maps of the L-band radiation by means of interferometry in order to obtain a reasonable angular resolution.One key aspect in interferometric imaging is that the average brightness temperature level of each measured image is obtained from an independent measurement. In SMOS, there are three special Noise Injection Radiometers (NIR units) for this purpose. Obviously, accuracy of these NIR units, or reference radiometers, plays an important role in the performance of the whole MIRAS.Performance of any radiometer is often characterized by thermal and temporal stability of its measurements. Especially, since the scientific mission requirements of the SMOS mission rely on averaging data from several orbits, the stability of the measurements play an enhanced role. Throughout the SMOS mission the performance of the imaging system, as well as the performance of the NIR units, has been monitored with measurements of some well-behaving natural targets. Such targets have been the sky (which is also the only external calibration reference of MIRAS), Antarctica, and Pacific Ocean. These targets have been analyzed by Level 1 and Level 2 teams e.g. in terms of temporal stability, spatial ripple, and thermal stability of measurements.One key result from the investigations of the three years is that there has been a clear change in thermal and electrical properties of the MIRAS hardware during the first months of the mission, and that this change has settled or at least decreased along with the mission. This change has been attributed to changes in thermal and electrical conductivity in the ground plane cover around the antennas. By possibly affecting to the attenuation, efficiency, matching, and antenna patterns of individual units, these changes restrict the performance of MIRAS and the NIR units.Within the NIR units, different models have been proposed to compensate these effects. Mainly, the methods are based on relating some NIR unit parameter with the physical temperature of the antenna patch. These models have been clearly enhancing the performance of the units; however, room for further improvements still exists in this field. In this contribution we review the front-end models of the NIR units that have been developed during the three-year life time of SMOS. We explain measurements which are used to assess the different models and the key results which yield in the development of them. Especially, we show measurements of the validation targets (sky, Antarctica and Pacific Ocean) with different NIR front-end models. From these results we conclude the current radiometric stability of the measurements of SMOS, and point out possible new ways to further improve the performance.