Skip to main content
Back

Jina AI's jina-embeddings-v3 Beats OpenAI and Cohere With Just 570M Parameters

Jina AI's 570M parameter embedding model combines task-specific LoRA adapters, 8192 token context, and Matryoshka truncation across 89 languages.

  • Jina released jina-embeddings-v3, a 570M parameter multilingual embedding model with 8192 token context.
  • Ships with five task-specific LoRA adapters for retrieval, clustering, classification, and text matching.
  • Matryoshka Representation Learning lets you truncate embeddings from 1024 down to 32 dimensions.
  • Beats OpenAI and Cohere embeddings on English MTEB, beats multilingual-e5-large-instruct across multilingual tasks.
  • Built on jina-XLM-RoBERTa with 24 layers, RoPE positions, and adapters adding under 3% parameters.
  • Weights under CC BY-NC 4.0; commercial use via Jina API, AWS, or Azure.

Jina AI has released jina-embeddings-v3, a compact multilingual embedding model built around a clever architectural trick. Instead of one generic embedding head, it ships with five swappable LoRA adapters, each tuned for a different downstream task. The base model has 570 million parameters, handles context lengths up to 8,192 tokens, and posts state-of-the-art numbers on multilingual and long-context retrieval.

That combination of small footprint, long context, and task specialization is why it has passed 2.5 million downloads on Hugging Face and become a default building block for RAG pipelines that need to handle more than English.

Five adapters, one base model

Most embedding models force a choice between a single generic representation and a heavier instruction-tuned model that reads a natural language prompt before encoding. Jina's approach swaps in task-specific LoRA adapters, giving you specialized embeddings without prompt engineering.

You select the behavior at inference time by naming the task. The available adapters:

  • retrieval.query for encoding user queries in asymmetric search
  • retrieval.passage for encoding documents in the corpus
  • separation for clustering and re-ranking
  • classification for classification heads
  • text-matching for symmetric similarity like STS