jina-embeddings-v5-omni - a jinaai Collection
Multimodal (text + image + video + audio) embedding models aligned with jina-embeddings-v5-text-*. Two sizes, four task variants each.
a jinaai Collection
jina-embeddings-v5-omni: Text-Geometry-Preserving Multimodal Embeddings via Frozen-Tower Composition
In this work, we introduce frozen-encoder model composition, a novel approach to multimodal embedding models. We build on the VLM-style architecture, in which non-text encoders are adapted to produce input for a language model, which in turn generates embeddings for all varieties of input. We present the result: the jina-embeddings-v5-omni suite, a pair of models that encode text, image, audio, and video input into a single semantic embedding space. Our method is to extend the two Jina Embeddings v5 Text models to support additional media by adding encoders for images and audio. The backbone text embedding models and the added non-text media encoders remain frozen. We only trained the connecting components, representing 0.35% of the total weights of the joint model. Training is therefore much more efficient than full-parameter retraining. Additionally, the language model remains effectively unaltered, producing exactly the same embeddings for text inputs as the Jina Embeddings v5 Text models. Our evaluations show that this approach produces results that are competitive with the state-of-the-art, yielding nearly equal performance to larger multimodal embedding models.
arXiv.orgFlorian Hönicke
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.
Architecture
v5-omni keeps the v5-text backbone completely frozen and adds pretrained vision and audio encoders connected through small trainable projectors:
- Vision: Qwen3.5 vision encoders (adapted from SigLIP2) with 2x2 spatial merge (4x token reduction). We freeze everything except the final projection layer (fc_vision_2), which we replace with a randomly initialized layer mapping into the text backbone's hidden dimension.
- Audio: Qwen2.5-Omni encoder (adapted from Whisper-large-v3). A single randomly initialized fc_audio layer projects the 1280-dimensional output into the text backbone.
- Video: Handled as a sequence of visual frames, optionally preceded by an extracted audio segment.
The model inherits v5-text's four task-specific LoRA adapters (retrieval, text-matching, classification, clustering) and trains separate projector weights for each task variant. The architecture is fully modular: text-only deployment loads no vision or audio weights (identical to v5-text footprint), image-only skips audio, full omni loads everything.
Getting Started
Elasticsearch (Elastic Inference Service)
If you are already using jina-embeddings-v5-text in Elasticsearch, your existing text indexes work with v5-omni out of the box. The omni models produce identical embeddings for text inputs as v5-text - same input, same vector, byte-for-byte. You do not need to re-embed or rebuild any text index. To start searching images, audio, and video alongside your existing text data, simply create a new index with v5-omni and ingest your multimodal content into it.
Create a semantic_text index with v5-omni as the inference endpoint. EIS automatically selects the correct LoRA adapter for indexing and retrieval:
Ingest text, images (as base64 data URIs), audio, and video into the same field, the same index:
Search across all modalities with a single text query:
Jina Embedding API
Hugging Face
Training
The core idea is frozen-encoder model composition: take a strong text embedding model, add pretrained vision and audio encoders, connect them with small trainable projectors, and freeze everything except those projectors. Only 0.35% of total weights are trained, which gives us three properties: (1) text identity preservation - the backbone is unmodified, same input produces identical output; (2) training efficiency - projector-only training is 1.8-3.9x faster with 42-64% less GPU memory; (3) modularity - towers can be loaded independently.
v5-omni inherits Matryoshka dimension support from v5-text. Image and audio embeddings preserve most quality under truncation, while video degrades more at small dimensions.
Conclusion
The conventional wisdom says multimodal embeddings require training the entire model end-to-end. We disagree. v5-omni freezes the text backbone, trains 0.35% of weights, and matches models 5-7x its size. The lesson: composition beats retraining. A strong text encoder is the hardest part – once you have it, bolting on vision and audio via lightweight projectors is almost free.
This matters for production. Your existing v5-text indexes is untouched. Same query, same vector, byte-for-byte. You just gained image, audio, and video search without re-embedding a single document. That is the real unlock multimodal retrieval as a drop-in upgrade, not a migration project.
jina-embeddings-v5-omni-small is the best-performing open-weight omni embedding model under 2B parameters. jina-embeddings-v5-omni-nano does it at 0.9B. Both available now on Hugging Face, Jina Search Foundation API, and as a native inference endpoint in Elasticsearch.
