What is a multi-vector model?

A multi-vector model keeps one small vector per token instead of compressing into a single vector. It scores queries by matching each query token to its best document token.

How long does training take?

The example finetuned model was trained in 14.5 hours on a single RTX 3090.

Should I start from an existing checkpoint or a fresh projection?

Start from a pre-supervised -unsupervised checkpoint if available, as they adapt better than finished checkpoints. A fresh projection on a strong retrieval-pretrained backbone is a close runner-up.

Large Language ModelsOpen-Source AIHugging Face Blog1h agoAug 27, 2026, 01:01 JST2 min read

Sentence Transformers v6.0 adds MultiVectorEncoder training

Send on LINE

Key takeaway

Sentence Transformers now supports training multi-vector retrieval models. Finetuning on your own domain can outperform general-purpose retrievers.
Even a fresh projection on a strong backbone reaches within 0.03 of existing checkpoints.
Training can run in hours on a single consumer GPU.

3 Key Points

What happened
Sentence Transformers v6.0 adds a fourth model type, MultiVectorEncoder, for ColBERT-style late-interaction retrieval. It now supports finetuning these multi-vector models, and the method can also train new ones from scratch.
Why it matters
Finetuning multi-vector models significantly improves retrieval on your domain. On a medical evaluation with 941-token passages, truncation from many pretrained models costs up to 0.24 NDCG@10. Also, the -unsupervised checkpoints adapt far better than finished siblings, so starting from a pre-supervised checkpoint is recommended.
What to watch
A finetuned model (multi-vector-encoder/mLateOn-medical) trained in 14.5 hours on a single RTX 3090 outperformed every general-purpose retrieval model on the medical evaluation. You can try the training pipeline with `pip install -U "sentence-transformers[train]"`.

Ask the AI about this article →

Summaries like this, in your inbox every morning.

Get news like this every morning — free

Context & Analysis

This update addresses a common pain point in retrieval: most released models truncate long documents, costing up to 0.24 NDCG@10 on medical passages averaging 941 tokens. Training your own multi-vector model lets you set the document length your data needs, avoiding this silent loss.

The author's experiments show that -unsupervised checkpoints, which have not been finetuned on general retrieval, adapt to new domains better than finished siblings. This is because they carry late-interaction structure without the general-purpose tuning that domain training must undo. A fresh projection on a strong backbone like gte-modernbert-base can nearly match existing checkpoints, making it a viable fallback.

The practical impact is that teams can now build domain-specific retrieval models in hours on a single GPU, without needing extensive resources. The provided training pipeline simplifies the process, and the evaluation demonstrates that even a modestly trained model can beat all general-purpose retrievers in a specific domain.

FAQ

What is a multi-vector model?: A multi-vector model keeps one small vector per token instead of compressing into a single vector. It scores queries by matching each query token to its best document token.
How long does training take?: The example finetuned model was trained in 14.5 hours on a single RTX 3090.
Should I start from an existing checkpoint or a fresh projection?: Start from a pre-supervised -unsupervised checkpoint if available, as they adapt better than finished checkpoints. A fresh projection on a strong retrieval-pretrained backbone is a close runner-up.

Hugging Face BlogRead Original Article