Sentence Transformers v6: ColBERT-Style Multi-Vector Models Fully Supported

Melvin Vivas · X post · 2026-08-20 · Open on X

Topics: Embeddings & Vector Databases, Retrieval-Augmented Generation (RAG) · Level: intermediate

Summary

Sentence Transformers v6.0 has been released. It adds MultiVectorEncoder, which makes ColBERT-style late-interaction models a full model type alongside dense, sparse and reranker models. Training, inference and interpretation are all supported.

Key points

• Sentence Transformers v6.0 adds the MultiVectorEncoder model type.
• ColBERT-style late interaction stores one vector per token instead of one vector per document.
• Multi-vector models can now be trained, run and interpreted inside the library.
• The library now covers four model families: dense, sparse, reranker (cross-encoder) and multi-vector.

Resources mentioned

• Sentence Transformers · tool · sbert.net · free
  Open-source Python library for training and running embedding, sparse, reranker and (from v6) multi-vector models.
• ColBERT · paper · arxiv.org · free
  Late-interaction retrieval model that scores documents by matching their token embeddings against the query's token embeddings.

Try this

• Upgrade to Sentence Transformers v6 and try MultiVectorEncoder for late-interaction retrieval.
• Compare dense, sparse and ColBERT-style multi-vector retrieval on your own RAG dataset.