Jina ColBERT v2
Jina ColBERT v2 uses late interaction for multilingual search. It encodes tokens separately so a query can match detailed parts of a document rather than only one pooled document vector.
How the model works
The model supports an 8,192-token input and Matryoshka token vectors. Its main package uses 128 dimensions; Jina also documents 96- and 64-dimensional variants. These widths trade index size against the information kept in each token representation. They are variants of the same v2 release.Jina ColBERT v2 model details
Use and access
The official card includes Sentence Transformers, PyLate, RAGatouille and ColBERT integration paths. Downloaded weights use CC BY-NC 4.0. Compared with v1, v2 extends beyond English and adds adjustable vector width. A late-interaction index must retain the token-level structure needed for scoring; it is not interchangeable with a single-vector index.Jina ColBERT v2 model detailsJina ColBERT v2 technical overview
Specifications & access
| Detail | Value and conditions |
|---|---|
| Launch date | |
| Access | API, Downloadable |
| Model identifiers | jina-colbert-v2jinaai/jina-colbert-v2jina-colbert-v2-96jina-colbert-v2-64 |
| Context window | 8,192 tokens Maximum input sequence in the model card; serving limits can be lower. Jina ColBERT v2 model details |
| Embedding dimensions | 128, 96 or 64 per tokenJina ColBERT v2 model details |
| Weight license | CC BY-NC 4.0Jina ColBERT v2 model details |
Related models
- Jina Embeddings v4
Jina AI · Jina Embeddings v4
A visual-document encoder with both single-vector and token-vector retrieval.
Launch date API · Downloadable - Jina ColBERT v1 English
Jina AI · Jina ColBERT
An English token-level retriever built for longer documents.
Launch date Downloadable