MoLFormer

MoLFormer is IBM Research's chemical language model, pre-trained on SMILES string representations of up to 1.1 billion molecules from the ZINC and PubChem datasets. It uses masked language modeling with a linear-attention Transformer and rotary embeddings for efficiency at that scale.

The foundation model is fine-tuned for downstream molecular property prediction tasks, and its encodings have been shown to recover molecular similarity and reflect interatomic spatial distances. It is widely used in drug-discovery embedding pipelines. Release timing below is era-level; consult the paper for the exact publication date.