Observations from the Multidisciplinary drifting Observatory for the Study of Arctic Climate (MOSAiC) were used to evaluate the Coupled Arctic Forecast System (CAFS) model’s ability to simulate the atmospheric boundary layer (ABL) structure in the central Arctic. MOSAiC observations of the lower atmosphere from radiosondes, downwelling longwave radiation (LWD) from a pyranometer, and near-surface wind conditions from a meteorological tower were compared to 6-hourly CAFS output. A self-organizing map (SOM) analysis reveals that CAFS reproduces the range of stability structures identified by the SOM trained with MOSAiC observations of virtual potential temperature (θv) profiles, but not necessarily with the correct frequency or at the correct time. Additionally, the wind speed profiles corresponding to a particular θv profile are not consistent between CAFS and the observations. When categorizing profiles by static stability, it was revealed that CAFS simulates all observed stability regimes, but overrepresents the frequency of near-surface strong stability, and underrepresents the frequency of strong stability between the top of the ABL and 1 km. The 10 m wind speeds corresponding to each stability regime consistently have larger values in CAFS versus observed, and this offset increases with decreasing stability. Whether LWD is over or underestimated in CAFS is dependent on stability regime. Both variables are most greatly overestimated in spring, leading to the largest near-surface θv bias, and the greatest underrepresentation of strong stability in spring. The results of this article serve to highlight the positive aspects of CAFS for representing the ABL and reveal impacts of misrepresentations of physical processes dictating energy, moisture, and momentum transfer in the lower troposphere on the simulation of central Arctic ABL structure and stability. This highlights potential areas for improvement in CAFS and other numerical weather prediction models. The SOM-based analysis especially provides a unique opportunity for process-based model evaluation.