Jina Embeddings v3
Multilingual text embedding model with task-specific adapters
Jina AI released Jina Embeddings v3 in September 2024. This 570-million-parameter model creates vector representations of text for search and classification tasks. It supports long contexts and multiple languages using specialised adapters.
- Contains 570 million parameters
- Supports context lengths up to 8192 tokens
- Uses task-specific LoRA adapters
- Allows dimension reduction via Matryoshka Learning
- Outperforms OpenAI embeddings on English MTEB tasks
What it is
Jina Embeddings v3 is a text embedding model developed by Jina AI. It converts text into numerical vectors that computers can compare for similarity. The model contains 570 million parameters and handles input sequences up to 8192 tokens long. It is designed for multilingual applications and supports various downstream tasks such as retrieval, clustering, and classification through specific adapters.
How it came about
The team at Jina AI introduced this model in September 2024 via a technical paper on arXiv. The authors include Saba Sturua, Isabelle Mohr, and Han Xiao among others. This release followed earlier versions in the Jina embeddings series, aiming to improve performance on both English and multilingual benchmarks compared to prior iterations and competitor models.
Inside the model
The architecture uses an encoder-only transformer design. A key feature is the inclusion of task-specific Low-Rank Adaptation (LoRA) adapters. These allow the same base model to generate high-quality embeddings for different purposes like query-document retrieval or text matching. The model also employs Matryoshka Representation Learning, which lets users reduce the output dimension from 1024 down to 32 without significant loss in performance.
What it can do
Evaluations on the MTEB benchmark indicate strong performance. The model reportedly outperforms proprietary embeddings from OpenAI and Cohere on English tasks. It also achieves superior results compared to multilingual-e5-large-instruct across all multilingual tasks tested. The system supports flexible output dimensions and maintains accuracy even when compressed to lower sizes.
How to use it
Jina AI provides the model weights openly. Developers can download the model from platforms like Hugging Face. The open nature of the release allows for local deployment and integration into custom search or classification pipelines without relying on external API calls for inference.
Questions people ask about Jina Embeddings v3
What is Jina Embeddings v3 used for?
It creates vector representations of text for tasks like semantic search, document retrieval, clustering, and classification. The model helps systems understand the meaning of text rather than just matching keywords.
How many languages does it support?
The model is designed for multilingual use. Evaluations show it performs well across various languages, outperforming other multilingual models like multilingual-e5-large-instruct on standard benchmarks.
Can I change the size of the output vectors?
Yes. Using Matryoshka Representation Learning, you can reduce the default output dimension of 1024 down to as low as 32. This flexibility helps save storage space while maintaining performance quality.
Good reading and sources
There is no English Wikipedia article on Jina Embeddings v3 yet, so the facts on this page come from the sources listed above: huggingface.co/jinaai/jina-embeddings-v3, arxiv.org/abs/2409.10173. Some are the subject's own pages, so its claims about itself are reported as its claims. EveryLLM wrote the text in its own words, with AI assistance, and compares it with the sources to make sure no passage is copied.
It may contain mistakes, and figures change over time. Nothing here is investment or buying advice. If something is wrong, tell us. Last refreshed 2026-11-14.