Jina Embeddings 5 Omni Small

What it does

Jina Embeddings 5 Omni Small extends Jina 5 Text Small with vision and audio encoders. The official release describes a shared vector space for text, images, audio, video and PDFs, with 1,024-dimensional vectors and a 32,768-token context.

Jina says its frozen text backbone produces identical text embeddings to the corresponding Jina 5 Text model. Compatibility assumes matching model variant, task and output parameters. This can let an existing Jina Text Small collection expand to other media without regenerating its text vectors. Video uses visual frames, optionally with extracted audio; the release reports weaker temporal/video retrieval than some larger baselines.

Where it helps

When to choose it

Evaluate Small for broader multilingual documents and richer media. Measure video retrieval against a video-focused model on your own clips.

Media adapter implemented in AIRuntime; commercial activation, provider credentials and verified pricing are required before calls are enabled. Inline payloads are bounded to 1 MiB per request; remote media URLs and large-file processing are not exposed. Always use a compatible model, task and dimension for indexed content and queries.