The Sentence Transformers library now supports training and finetuning multi-vector embedding models. This update enables ColBERT-style late interaction, which improves retrieval accuracy over single-vector representations. Developers can now optimize complex document representations using standard training loops.