We introduce ModernMolBERT, an encoder-only transformer model for small-molecule representation learning. ModernMolBERT adapts the ModernBERT architecture to molecular language modelling and pairs it with a SELFIES-aware Atom Pair Encoding (APE) tokeniser. Both released variants (small and base) are pre-trained with masked language modelling on approximately 2.4M unique SELFIES strings from ChEMBL 36 and evaluated as frozen molecular embedders on 25 binary classification benchmarks. ModernMolBERT-base attains a mean ROC-AUC of 77.9 (reported on a 0 to 100 scale), competitive with a strong fingerprint baseline and improving on comparable pre-trained string encoders by roughly five points. ModernMolBERT provides a compact, reproducible, open-source foundation for SELFIES-based molecular representation learning; we release the model weights, the APE tokeniser, and training code.