Recent breakthroughs in foundation models have revolutionized the science domain with their promising generalization performance to solve challenging open questions. In chemistry and biology, the textual data enriches comprehensive knowledge about the molecule's functionalities, thus serving as a complementary modality to the chemical structures. However, existing multi-modal foundation models mainly focus on 2D topology rather than 3D geometry. To handle this issue, we construct a large-scale 3D structure-text dataset with conformations calculated by semi-empirical quantum methods. Then we propose MoleculeSTM-3D, a geometry-text multi-modal foundation model to align the two modalities through contrastive learning. For downstream, we apply MoleculeSTM-3D to the reactivity-oriented molecule editing task. Our empirical results demonstrate that it achieves a 9.00% higher editing success rate and significantly reduces invalid molecule generation by 10.07% compared to baseline methods. Such preliminary results reveal the potential of utilizing MoleculeSTM-3D for solving more challenging tasks.