Jina Embeddings V3

jina-embeddings-v3 is a multilingual multi-task text embedding model that generates high-quality vector representations for retrieval, clustering, classification, and text matching tasks. Based on the Jina-XLM-RoBERTa architecture with 570 million parameters, it supports context lengths up to 8192 tokens via Rotary Position Embeddings (RoPE) and uses five task-specific LoRA adapters to produce optimized embeddings.

Key capabilities

• Task-specific LoRA adapters: retrieval.query, retrieval.passage, separation, classification, and text-matching.
• Matryoshka embeddings: Flexible output dimensions from 1024 down to 32 without degrading performance.
• Multilingual support: Optimized for 30 languages (Arabic, Bengali, Chinese, Danish, Dutch, English, Finnish, French, Georgian, German, Greek, Hindi, Indonesian, Italian, Japanese, Korean, Latvian, Norwegian, Polish, Portuguese, Romanian, Russian, Slovak, Spanish, Swedish, Thai, Turkish, Ukrainian, Urdu, Vietnamese); foundation model covers 100 languages.

Benchmark results

On the MTEB benchmark, jina-embeddings-v3 outperforms OpenAI text-embedding-3-large and Cohere embed-multilingual-v3.0 on English tasks, and surpasses multilingual-e5-large-instruct on all multilingual tasks. It is the best-performing multilingual model among those with fewer than 1 billion parameters (ranked 2nd on the English MTEB leaderboard as of September 2024). On six long-document retrieval tasks (LongEmbed), it shows significant improvements over models using fixed or ALiBi positional embeddings.

Efficiency

Compared to e5-mistral-7b-instruct (7.1B parameters, 4096 output dimension), jina-embeddings-v3 is 12× smaller and uses 4× smaller output dimensions while achieving comparable or better English MTEB scores, making it suitable for production and edge deployment.