Jina AIJina Embeddings v3

Jina Embeddings v3

Jina Embeddings v3 is a multilingual encoder for retrieval, classification and text similarity. Its Jina-XLM-RoBERTa design supports 8,192-token inputs and task-specific LoRA adapters.

How the model works

Five adapters cover retrieval queries, retrieval passages, separation, classification and text matching. Mean pooling creates vectors, and Matryoshka training permits dimensions from 32 to 1,024. Jina says it specifically tuned 30 languages, while the underlying model supports a broader language set. These are different claims and should not be treated as one coverage count.Jina Embeddings v3 model details

Use and access

The model card records a correction to vector truncation: truncating after normalization could produce non-normalized shortened vectors. Existing indexes can depend on the code revision used to generate them. Jina documents that revision for consistency. The model remains downloadable under CC BY-NC 4.0 and has API access, although newer v5 text models are available.Jina Embeddings v3 model detailsJina Embeddings v3 technical overview

Specifications & access

Jina Embeddings v3 specifications
DetailValue and conditions
Launch date
AccessAPI, Downloadable
Model identifiersjina-embeddings-v3jinaai/jina-embeddings-v3jina-embeddings-v3-hf
Context window8,192 tokens

Maximum input sequence in the model card; serving limits can be lower.

Jina Embeddings v3 model details
ParametersApproximately 570 millionJina Embeddings v3 model details
Embedding dimensions1,024; Matryoshka sizes down to 32Jina Embeddings v3 model details
Weight licenseCC BY-NC 4.0Jina Embeddings v3 model details
Base modelJina-XLM-RoBERTaJina Embeddings v3 model details

Related models

  • A multilingual text encoder with four task adapters and compact vector options.

    Launch date API · Downloadable
  • Jina CLIP v2

    Jina AI · Jina CLIP

    A multilingual text-and-image encoder with adjustable vector dimensions.

    Launch date API · Downloadable