This study investigates the accuracy of computational fluid dynamics (CFD) models to predict heat transfer in turbulent separated flows at low Reynolds numbers. This article will focus on flow in a staggered tube bank, while its companion articular will focus on a square prism (cylinder) in cross flow. Experimental data for both local heat transfer and velocity profiles are available for these cases and have been used extensively in the literature to evaluate various CFD methods. Six unsteady models were used and the results show that the unsteady shear stress transport (SST) model provided good overall accuracy relative to the mean Nusselt number for both cases. However, the SST model failed to accurately predict local variations. The partially averaged Navier–Stokes (PANS) variant of the SST model did show a marked improvement over the baseline SST model. The dynamic Smagorinsky large eddy simulation (LES) showed a much-improved fidelity to the local Nusselt number but unpredicted the actual values. The computational cost for the LES model was significant and it was found that the computationally expensive models with higher degrees of resolved turbulence did not necessarily return better results. Finally, the pressure drop results for the six models were scaled to predict the mean Nusselt number with the generalized Leveque method and were found to be very accurate. This method should prove useful to predict heat transfer performance with computationally less expensive cold flow results.