We introduce Zatom-1, a multimodal foundation model for 3D small molecules and materials that generates realistic molecules, thermodynamically stable materials, and valid metal-organic frameworks while achieving competitive performance for PoseBusters and LeMat-GenBench. By pretraining across multiple atomistic modalities with a self-supervised generative flow matching objective, we demonstrate the first case of positive predictive transfer learning where modeling materials enhances molecular property prediction accuracy for downstream tasks like QM9 and Matbench. Zatom-1 is powered by an approximately equivariant transformer architecture that scales effectively in ambient space and outperforms existing equivariant and latent generative models.