Introducing Jina Embeddings v4, a cutting-edge 3.8 billion parameter multilingual and multimodal embedding model, now available on Amazon SageMaker. This model aligns text and image inputs in a shared embedding space and achieves state-of-the-art retrieval performance on multimodal and multilingual tasks across MTEB, MMTEB, CoIR, LongEmbed, STS, Jina-VDR, CLIP, and ViDoRe benchmarks. Deploy seamlessly within your AWS infrastructure to support scalable inference and simplified MLOps workflows.

Jina Embeddings v4 features a novel architecture that supports both single-vector and multi-vector embeddings through late interaction, offering greater flexibility than traditional dual-encoder approaches. It includes a frozen backbone and multiple task-specific LoRA adapters, enabling optimized performance for use cases including query-document retrieval, semantic matching, and code search. The model processes visually rich inputs such as tables, charts, diagrams, and images through a unified embedding pathway, allowing for consistent cross-modal retrieval. With support for inputs up to 32,768 tokens and training data spanning 29 languages, it is well-suited for long-form document and multilingual search.