jina embeddings v5 omni small

jina-embeddings-v5-omni-small is a multimodal embedding model that accepts text, images, video, and audio and produces embeddings in a shared vector space aligned with the text-only jinaai/jina-embeddings-v5-text-small, enabling index-with-text and query-with-any-modality without reindexing.

Architecture

The model uses the GELATO (Geometry-preserving Embeddings via Locked Aligned TOwers) approach. Non-text encoders are adapted to feed a frozen language model that generates embeddings for all input types. Only the connecting components are trained—0.35% of total weights. The text backbone remains bit-identical to the text-only Jina Embeddings v5 model. The vision encoder is a fine-tuned SigLIP2 So400m (from Qwen3.5-2B) and the audio encoder is Whisper-large-v3 (from Qwen2.5-Omni-7B).

Benchmark Performance

• Best performance on image benchmarks of any model in the ~1 billion parameter range; only models with 3–30 times as many parameters beat it.
• State-of-the-art multilingual visual understanding, outperforming models up to 20× larger.
• Best-in-size-class audio embeddings; only models with double or more parameters perform better on standard benchmarks.

The model supports retrieval, classification, clustering, and text-matching tasks via separate adapters. Embedding dimension is 1024, with Matryoshka dimensions from 32 to 1024. Maximum sequence length is 32768 tokens.