In this study, the ability of a turbulent kinetic energy (TKE)–based boundary layer scheme to reproduce the rapid evolution of the planetary boundary layer (PBL) observed during two clear convective days is examined together with the impact of including nonlocal features in the boundary layer scheme. The two cases are chosen from the Montreal-96 Experiment on Regional Mixing and Ozone (MERMOZ): one is characterized by strong buoyancy, a strong capping inversion, and weak vertical wind shear; the other displays moderate buoyancy, a weaker subsidence inversion, and significant wind shear near the PBL top. With the original local version of the turbulence scheme, the model reproduces the vertical structures and turbulent quantities observed in the well-developed boundary layer for the first case. For the second case, the model fails to reproduce the rapid evolution of the boundary layer even though the TKE and sensible heat fluxes are greatly overpredicted. Some nonlocal aspects of the turbulence scheme are tested for these two cases. Inclusion of nonlocal (countergradient) terms in the vertical diffusivity equation has little impact on the simulated PBL. In contrast, alternative formulations of the turbulent length scales that follow the strategy proposed by Bougeault and Lacarrère have a greater influence. With the new turbulent lengths, entrainment at the top of the boundary layer is enhanced so that the depth of the well-mixed layer is much larger compared to that of the local simulations even though the turbulent sensible heat fluxes are smaller. Comparison with observations reveals, however, that the inclusion of these modifications does not improve all aspects of the simulation. To improve the performance and reduce somewhat the arbitrariness in the Bougeault–Lacarrère technique, a relationship between the two turbulent length scales (mixing and dissipation) used in the turbulence scheme is proposed. It is shown that, in addition to reducing the sensitivity of the results to the particular formulations, the simulated boundary layer agrees better with observations.