Jina-CLIP v2

Overview

Jina-CLIP v2 is a state-of-the-art multimodal embedding model that combines text and image understanding in a single unified model. It represents a significant improvement over v1 with enhanced multilingual capabilities and higher resolution image processing.

Architecture

The model combines two specialized encoders:

• Text Encoder: Jina XLM-RoBERTa (561M parameters)
• Vision Encoder: EVA02-L14 (304M parameters)
• Total Parameters: 865M

Key Features

• Multilingual Support: Supports 89 languages for text-image retrieval with up to 4% improvement over comparable models
• High Resolution: Processes 512x512 images, a significant upgrade from v1's 224x224 resolution
• Matryoshka Representations: Allows truncating output dimensions from 1024 to 64 while maintaining 99% performance
• State-of-the-Art Performance: Achieves 98.0% accuracy on Flickr30k image-to-text retrieval
• Flexible Deployment: Available via Jina Embeddings API, AWS, Azure, and GCP

Performance

Even aggressive 75% dimensional reduction maintained over 99% performance across text, image, and cross-modal tasks. The model shows 3% performance improvement over v1 in both text-image and text-text retrieval tasks.

Use Cases

• Cross-modal search (text-to-image, image-to-text)
• Multilingual image retrieval
• Visual question answering
• Content-based recommendation systems
• Multimodal RAG applications

Pricing

Available through Jina Embeddings API with commercial licensing. Also available on cloud marketplaces (AWS, Azure, GCP) with usage-based pricing.