Sentence Transformers added a fourth model type, MultiVectorEncoder, in its v6.0 release to support ColBERT-style late interaction retrieval, Hugging Face announced on August 18, 2026. The update allows users to load native checkpoints, Stanford-NLP ColBERT models, PyLate checkpoints, and colpali-engine models for visual document retrieval through its standard interface.
Standard dense embedding models compress an entire passage into a single fixed-size vector of 384, 768, or 1024 dimensions, calculating similarity with a single dot product. In contrast, multi-vector models keep one contextualized vector per token—typically projected down to 128 dimensions—and defer token comparison until scoring time. The system calculates similarity using the MaxSim operator, which matches each query token against the most similar document token and sums the resulting scores.
Token Storage and Indexing
The token-level approach improves retrieval on multi-requirement queries and exact identifier searches, but increases index size. In tests on 4,874 Natural Questions passages, Hugging Face reported that the lightonai/LateOn model generated 608,414 token vectors, requiring 311.5 MB of uncompressed float32 storage compared to 7.5 MB for the single-vector all-MiniLM-L6-v2 model. When indexed with LightOn's fast-plaid library using centroid and quantized residual compression, that same index required 92 MB.
Because queries and documents use separate marker prefixes, length caps, and scoring masks, the architecture separates ingestion into encode_query and encode_document methods. For setups where maintaining a late-interaction index is impractical, the model can operate as a second-stage reranker over top candidate results retrieved by a conventional dense bi-encoder.
Integration and Dependencies
Several external vector databases support multi-vector indexing and MaxSim scoring natively, including Qdrant since version 1.10, Weaviate since version 1.29, LanceDB since version 0.15.0, Milvus since version 2.6.4, Vespa, and VectorChord. OpenSearch and Elasticsearch support candidate rescoring with MaxSim, while turbopuffer has indexing in private beta.
Sentence Transformers v6.0 requires transformers v5.x, torch 2.2 or higher, and huggingface-hub v1.x. For visual document retrieval models that score page images directly without optical character recognition, installations require the optional image dependencies package.
