The study evaluates OpenFOAM's adaptive mesh refinement (AMR) capabilities and accuracy when applied to numerical simulations of turbulent buoyant flows. To this purpose, large eddy simulations of Sandia's turbulent helium plume test case are considered, a scenario with similar characteristics (in terms of air entrainment and vortex shedding) to those encountered in large-scale fires. Comparison is made of the relative accuracy (in terms of both first and second order statistics), the predicted puffing frequencies and the computing times of numerical simulations using AMR and static meshes. Additionally, a sensitivity study on different AMR parameters is conducted and an evaluation of the performance of the dynamic Smagorinsky model when combined with AMR is made. Overall, the predictions of AMR are very satisfactory, in terms of both accuracy and computing times, compared to the use of static meshes of different sizes. Nevertheless, it is observed that a careful choice of a static mesh can result in equally accurate predictions with comparable, or even smaller, computing times than AMR for the case at hand. Additionally, the use of AMR slightly modified the entrainment characteristics in the near-field region of the plume and (significantly) altered some of the, dynamically determined, turbulence model parameters.