CheMeleon is an O(10M)-parameter directed message-passing neural network (D-MPNN) foundation model for molecular property prediction, pre-trained to regress classical Mordred molecular descriptors from molecular graphs; it is consumed either as a frozen 2048-dimensional molecular fingerprint generator or as the message-passing initialization for a fine-tuned ChemProp model.