Jina Embeddings V4
Jina Embeddings V4 is a universal embedding model designed to handle multiple modalities and languages within a single unified embedding space. Unlike single-modality embeddings that process only text or only images, V4 accepts text, images, and other content types as input, converting them into comparable vector representations. This multimodal approach enables retrieval systems to perform cross-modal semantic search, where a user query in one modality can retrieve relevant results from another modality.
The model is optimized for retrieval-augmented generation (RAG) and semantic search applications. By processing diverse content types through a shared embedding space, V4 supports more flexible information retrieval workflows where text searches can surface relevant images and vice versa. The multilingual capabilities extend this functionality across different languages, allowing the model to handle global datasets without separate language-specific models.
Jina Embeddings V4 represents an evolution in universal embedding models, addressing the practical need for systems that work across modalities and languages simultaneously. Rather than requiring multiple specialized models or preprocessing steps, V4 consolidates these capabilities into a single embedding framework, simplifying deployment in RAG pipelines and semantic search infrastructure.
Source Notes
- 2026-04-23: Anthropic · ▶ source