tech
jina-embeddings-v5-omni: Embeddings for Text, Image, Audio and Video
We are releasing jina-embeddings-v5-omni, extending our v5-text embedding models to images, audio, and video. Both models share the same frozen text backbone as v5-text, meaning text embeddings are identical - no index rebuild needed. jina-embeddings-v5-omni-small scores 53.93 on average across four modalities, matching LCO-7B (54.43) at 5.7x fewer parameters, while jina-embeddings-v5-omni-nano delivers competitive document retrieval at just 0.95B parameters.
TL;DR
- jina-embeddings-v5-omni models embed text, image, audio, and video into a single semantic space.
- They leverage a frozen text backbone from Jina Embeddings v5 Text models, ensuring identical text embeddings.
- Training involves only small connecting components (projectors), comprising 0.35% of total weights, leading to high efficiency.
- jina-embeddings-v5-omni-small (1.57B parameters) matches LCO-7B's average score with 5.7x fewer parameters.
- The architecture is modular, allowing for text-only, image-only, or full omni deployments.
- Existing text indexes using v5-text are compatible with v5-omni without re-embedding.
- The models show strong performance across modalities, with text, image, and audio being competitive, while video shows the most significant gap.
- jina-embeddings-v5-omni is available on Hugging Face, Jina Search Foundation API, and as a native inference endpoint in Elasticsearch.