How much more storage does a multi-vector index need?

Sentence Transformers v6.0 documents the tradeoff on 4,874 Natural Questions passages: the LateOn multi-vector index takes 311.5 MB raw, against 7.5 MB for a MiniLM dense index. Compression brings the multi-vector index down to about 92 MB. Token-level vectors buy retrieval quality and cost disk space, so budget storage before switching.

Which vector databases can serve these models?

MaxSim scoring is supported natively by Qdrant, Weaviate, Vespa, LanceDB and VectorChord, and the Sentence Transformers v6.0 post also points at fast-plaid for approximate indexing. That means a MultiVectorEncoder model can be served by an existing vector store rather than needing a custom retrieval engine written around late interaction.

What breaks when upgrading from Sentence Transformers v5?

Sentence Transformers v6.0 is a major version with real breaking changes. The similarity and similarity_pairwise attributes are now methods rather than properties, a bare list of message dicts counts as one conversation instead of a batch, custom modules need trust_remote_code=True, rerankers and similarity now score in float32, and the Normalize module moved.

Can it load existing ColBERT or PyLate models?

Yes. MultiVectorEncoder in Sentence Transformers v6.0 loads checkpoints from PyLate, Stanford-NLP ColBERT and colpali-engine directly, so late interaction models already published in those formats work without conversion. LightOn's earlier PyLate framework pioneered the approach, and the release was written with LightOn.

Does v6.0 work with older transformers versions?

No. Sentence Transformers v6.0 requires transformers 5.x, torch 2.2 or newer, numpy 1.24.0 or newer, scikit-learn 1.1.0 or newer, and huggingface-hub 1.x. Projects still pinned to transformers 4.x need to upgrade that dependency first, or stay on the Sentence Transformers v5.7 line.

Hugging Face · 2026-08-18 · major

Sentence Transformers v6.0 — ColBERT-style retrieval joins the library

Sentence Transformers v6.0 adds MultiVectorEncoder, a ColBERT-style late interaction model type that keeps one vector per token instead of one per text. Hugging Face built it with LightOn; it loads PyLate and Stanford ColBERT checkpoints.

Sentence Transformers multi-vector encoder banner with Hugging Face and LightOn branding

Late interaction retrieval is now a built-in Sentence Transformers model type, not a separate framework.

Quick facts

Version	6.0.0
Maker	Hugging Face, with LightOn
License	Apache-2.0
What's new	MultiVectorEncoder model type
Requires	transformers 5.x, torch 2.2+
Loads checkpoints from	PyLate, Stanford ColBERT, colpali-engine
Install	pip install sentence-transformers==6.0.0

Benchmarks

NanoBEIR (average NDCG@10)

LateOn (multi-vector)		0.6868
DenseOn (dense)		0.6764

source ↗

What is it?

MultiVectorEncoder is the new model type in Sentence Transformers v6.0, and it is the library's fourth after dense embeddings, sparse models and rerankers. It keeps one vector per token rather than squeezing a whole passage into a single vector. The same class loads PyLate, Stanford ColBERT and colpali-engine checkpoints, so existing late interaction models work through one API.

How does it work?

Scoring uses the MaxSim operator: for each query token, take the highest similarity against any document token, then add those maxima together. Because token-level detail survives, exact matches and synonyms are both preserved instead of being averaged into one summary vector. The cost is index size — token vectors take far more room than a single embedding per passage.

Why does it matter?

Teams building RAG and search have had to bolt on a second framework to try late interaction. Sentence Transformers v6.0 removes that step, so the same training, evaluation and inference code covers dense, sparse, reranking and multi-vector models. On NanoBEIR the multi-vector LateOn model scores 0.6868 NDCG@10 against 0.6764 for its dense counterpart.

Who is it for?

RAG and search engineers

Frequently asked questions

How much more storage does a multi-vector index need?: Sentence Transformers v6.0 documents the tradeoff on 4,874 Natural Questions passages: the LateOn multi-vector index takes 311.5 MB raw, against 7.5 MB for a MiniLM dense index. Compression brings the multi-vector index down to about 92 MB. Token-level vectors buy retrieval quality and cost disk space, so budget storage before switching.
Which vector databases can serve these models?: MaxSim scoring is supported natively by Qdrant, Weaviate, Vespa, LanceDB and VectorChord, and the Sentence Transformers v6.0 post also points at fast-plaid for approximate indexing. That means a MultiVectorEncoder model can be served by an existing vector store rather than needing a custom retrieval engine written around late interaction.
What breaks when upgrading from Sentence Transformers v5?: Sentence Transformers v6.0 is a major version with real breaking changes. The similarity and similarity_pairwise attributes are now methods rather than properties, a bare list of message dicts counts as one conversation instead of a batch, custom modules need trust_remote_code=True, rerankers and similarity now score in float32, and the Normalize module moved.
Can it load existing ColBERT or PyLate models?: Yes. MultiVectorEncoder in Sentence Transformers v6.0 loads checkpoints from PyLate, Stanford-NLP ColBERT and colpali-engine directly, so late interaction models already published in those formats work without conversion. LightOn's earlier PyLate framework pioneered the approach, and the release was written with LightOn.
Does v6.0 work with older transformers versions?: No. Sentence Transformers v6.0 requires transformers 5.x, torch 2.2 or newer, numpy 1.24.0 or newer, scikit-learn 1.1.0 or newer, and huggingface-hub 1.x. Projects still pinned to transformers 4.x need to upgrade that dependency first, or stay on the Sentence Transformers v5.7 line.

Try it

pip install sentence-transformers==6.0.0