jina-embeddings-v3

Overview

jina-embeddings-v3 is a frontier multilingual text embedding model with 570M parameters and 8192 token-length, outperforming the latest proprietary embeddings from OpenAI and Cohere on MTEB.

Key Features

• 570M Parameters: Balanced size for performance and efficiency
• Long Context: Supports up to 8192 tokens
• Task-Specific LoRA Adapters: Includes adapters for query-document retrieval, clustering, classification, and text matching
• Matryoshka Embeddings: Flexible dimension truncation (32, 64, 128, 256, 512, 768, 1024) without significant performance loss
• Multilingual: Supports 89 languages with strong multilingual performance

Performance

• Outperforms OpenAI and Cohere embeddings on English tasks
• Superior performance compared to multilingual-e5-large-instruct across all multilingual tasks
• Better results than e5-mistral-7b-instruct, which has 12x more parameters

Technical Details

• Default output dimension: 1024
• Supports dimension reduction to 32-768 with Matryoshka approach
• Available on Hugging Face: jinaai/jina-embeddings-v3
• Integration with Elastic Inference Service

Use Cases

• Semantic search
• Document retrieval
• Clustering
• Classification
• Text matching
• Cross-lingual search

Availability

• Hugging Face Hub
• Jina AI API
• Elastic Inference Service
• LlamaIndex integration

Pricing

Available through Jina AI's API with both free tier and commercial plans.