Sentence Transformers v6.0 ships MultiVectorEncoder for ColBERT-style retrieval
Sentence Transformers v6.0 adds MultiVectorEncoder, which keeps one vector per token instead of collapsing a document into one embedding, and loads any PyLate or Stanford ColBERT checkpoint directly. The edge is real but narrow: on NanoBEIR, LateOn averaged 0.6868 NDCG@10 to DenseOn's 0.6764 at the same model size, winning 9 of 13 datasets. The cost is index size — 4,874 passages became 608,414 token vectors at 311.5 MB, about 42x a dense index — so token pooling or PLAID compression is part of shipping it.
Source: huggingface.co ↗
A green sofa with the wrong legs ends up sitting close to the one you actually asked for.
Sentence Transformers v6.0 documentation
Why this matters
- → Late interaction retrieval preserves token-level matching, beating single-vector models on exact matches and m
- → Index size grows ~42x uncompressed but compresses to competitive territory with PLAID quantization
- → Unified API loads PyLate, ColBERT, and ColPali checkpoints interchangeably without format conversion
Dense gets competition