Computational materials discovery must often compromise accuracy to accommodate the massive quantity of calculations necessary. Pioneering studies in foundation or “universal” machine-learned interatomic potentials (MLIPs), such as MACE-MP0 and OMol25, provide avenues towards performing simulations with near quantum mechanical accuracy at more feasible costs. Until now, efforts have largely focused on molecular or inorganic structures and limited developments exist for periodic crystal systems such as metal-organic frameworks (MOFs). Realizing a generalizable MOF-based MLIP is challenging due to the breadth of chemical substructures that must be considered to fully encompass this diverse structure class. This work pursues this universal model objective by training on 1.85M density functional theory (DFT) configurations (PBE0-D4/def2-TZVP) sampled from ca. 1.4k MOF-derived structural building units (SBUs). These MLIPs were applied directly in the calculation of relevant vibrational properties and achieved commensurate performance— >20% RMSE improvements in phonon-derived thermal properties (i.e. entropy, free energy, and heat capacity) and total DOS/band frequencies—with more specialized potentials developed on smaller collections (i.e. hundreds) of crystal structures (e.g., MACE-MP-MOFs-v2). Additionally, their usage as foundation models for fine-tuning protocols which target DFT-quality guest adsorption simulations was probed using adsorbate-loaded configurations. \(CO_2\) and \(H_2\)O adsorption isotherms generated in this manner showed excellent agreement with experimental data even in cases with prior documented issues caused by classical force field-based parameters (e.g., group 13 elements, low pressure, etc.). These findings highlight the promise of applying lower-cost data generation on diverse molecular fragment libraries towards MLIP-accelerated studies of related periodic crystal structures.