Jina Embeddings v5 Nano - Omni

Jina Embeddings v5 Omni Nano is the compact multimodal member of the latest generation of Jina AI's open-weight embedding family. Built on a EuroBERT-210m text backbone with a SigLIP2 Base vision tower and a Whisper-large-v3 audio tower, its 1.04B parameters map text, images, video, audio, and PDF documents into a single shared embedding space, so you can index any modality and retrieve any other from one vector index. This enables cross-modal search, multimodal RAG, visual document retrieval over PDFs and scans, video moment retrieval, and audio semantic search without separate per-modality pipelines. Text embeddings are identical to jina-embeddings-v5-text-nano, so multimodal content drops into an existing v5-text-nano index with no reindexing. The model handles an 8,192-token text context, supports multilingual text across dozens of languages, and delivers competitive document retrieval at just over a billion parameters, outperforming similarly sized open-weight omni models on the multimodal Pareto frontier. Matryoshka Representation Learning lets you truncate embeddings from 768 down to 32 dimensions without retraining, trading storage cost for marginal recall loss. Four task-specific LoRA adapters (retrieval, text-matching, clustering, and classification) tune the same base model for different downstream workloads at request time.