Jina Embeddings v5 Omni

Jina AI's non-commercial embedding models for text, images, video and audio in one space, in small and nano sizes.

In short. Jina Embeddings v5 Omni turns text, pictures, video clips and audio recordings into number vectors in one shared space, so a text question can find a matching image, scan, clip or recording. It comes in a small and a nano size, and it is free for non-commercial use only.

Summary

Jina Embeddings v5 Omni is a pair of multimodal embedding models, small and nano, that extend Jina's v5 text models to images, video and audio. Jina describes the method, called GELATO, in arXiv 2605.08384, submitted on 8 May 2026, and announced the models with Elastic, which now includes Jina AI, on 11 May 2026. The text backbone and the added encoders stay frozen and only the connecting projectors were trained, about 0.35% of the joint model's weights according to the paper, so the text embeddings are identical to those of jina-embeddings-v5-text-small and -nano and an existing text index needs no rebuild. The image and video encoders are SigLIP2 encoders taken from Qwen3.5 models, a So400m one from Qwen3.5-2B for small and a Base one from Qwen3.5-0.8B for nano, and both models use the Whisper large v3 audio encoder extracted from Qwen2.5-Omni-7B. Each has four task adapters for retrieval, classification, clustering and text matching.

Small gives 1,024-dimension vectors from inputs of up to 32,768 tokens, nano gives 768 dimensions and 8,192 tokens, and both can be cut as short as 32 dimensions through Matryoshka truncation. Images are scaled to between 262,144 and 3,072,000 pixels and cut into 28 by 28 patches, video is sampled at up to 32 evenly spaced frames, and audio is split into 30-second segments of 750 tokens each, so a long video loses most of its detail and Jina says short clips work best. Jina reports that small has the best image-benchmark scores among models near one billion parameters, that nano beats models 10 to 60 times larger on visual document retrieval, and that on video moment search small scores 55.57 on Charades-STA and 58.93 on MomentSeeker, against 29.30 and 59.30 for ByteDance's closed Seed 1.6 embedding model. These are Jina's own results, shown mostly in charts. Jina lists finding specific videos from a description, image-to-image similarity, spoken-intent classification and mixed-media inputs as areas it is still improving.

The license is CC BY-NC 4.0, and the blog says commercial use needs a license from Elastic's sales team, while the card says to contact Jina. The BF16 weights are 3.25 GB for small and 1.97 GB for nano, plus about 40 MB and 14 MB per adapter, and you can load only the text, vision or audio parts to save memory. Loading needs trust_remote_code, Transformers 4.57 or later and PyTorch 2.5 or later, vLLM 0.20.1 is the validated serving route, and Jina also publishes pre-merged per-task repositories and GGUF and MLX builds. Queries and documents take "Query: " and "Document: " prefixes, and pooling is last-token. The sources disagree on size, with 1.66B for small in the blog, about 1.74B on the card and 1.63B counted by Hugging Face. In the 30 days to 11 November 2026, Hugging Face counted about 1.01 million downloads of the small repository, 48,000 of its pre-merged retrieval version and 6,700 of nano.

Good for

  • One index for text, images, scanned pages, video clips and audio, searched with text queries in many languages
  • Adding image, video and audio search to an existing Jina v5 text index without re-indexing
  • Multilingual speech retrieval and visual document retrieval from text queries
  • Locating the moment in a video that matches a text description

Watch out for

  • The license is non-commercial, so a product needs a separate license from Elastic or Jina
  • Videos are sampled at 32 frames and audio is cut into 30-second segments, so index short clips
  • Jina lists specific-video search, image-to-image similarity, spoken-intent classification and mixed inputs as weaker areas
  • The scores are Jina's own and mostly shown in charts, and each task adapter produces its own vectors, so do not mix tasks in one index

Performance claims come from the developer and have not been independently verified.

How much memory does it need?

Memory figures for each variant of Jina Embeddings v5 Omni. Image, video, speech and music pipelines also load a text encoder and a decoder, so the memory needed can be larger than the main checkpoint.

Variants Parameters Download size Memory needed Fits on
omni-nano (768-dimension vectors) BF16 990M 1.97 GB 2.27 GB Laptop or phone
omni-small (1,024-dimension vectors) BF16 1.63B 3.25 GB 3.74 GB Laptop or phone

Download size covers every component needed to run the model, such as the generator, text encoder and decoder. Memory needed is a stated or published measured minimum where one exists; otherwise it is the download size plus about 15%. Larger images, longer clips and higher resolutions need more.

The license in plain words

A Creative Commons license that lets anyone copy, share and adapt the work with credit, but only for non-commercial purposes. Some model cards name it without a version number, and the current version is 4.0.

  • Attribution or model name must be shown

More about CC BY-NC

See also

Frequently asked questions

What is Jina Embeddings v5 Omni?
Jina Embeddings v5 Omni turns text, pictures, video clips and audio recordings into number vectors in one shared space, so a text question can find a matching image, scan, clip or recording. It comes in a small and a nano size, and it is free for non-commercial use only.
Can I use Jina Embeddings v5 Omni commercially?
Not without separate permission. Jina Embeddings v5 Omni is released under the CC BY-NC, which allows research and personal use only.
Is Jina Embeddings v5 Omni open source?
Jina Embeddings v5 Omni is open-weight, not open source in the strict sense. You can download the model files, but the CC BY-NC adds conditions that the Open Source Initiative definition does not allow.
Can I run Jina Embeddings v5 Omni on my own computer?
Yes. The smallest variant, omni-nano (768-dimension vectors), needs about 2.27 GB of memory at 4-bit precision, which fits a modern laptop or a mid-range graphics card.
When was Jina Embeddings v5 Omni released?
Jina Embeddings v5 Omni was released by Jina AI on May 11, 2026.
How long a context window does Jina Embeddings v5 Omni have?
Jina Embeddings v5 Omni supports a context window of up to 32K tokens.

Official links

Sources

Specifications come from these primary sources. Benchmark claims, where mentioned, are reported by the developer.

Release date and facts last verified November 11, 2026.

Related models

Model Released Parameters Context License Smallest variant needs
Jina Embeddings v5 Jina AI 239M – 677M 32K CC BY-NC (Non-commercial) 0.552 GB Laptop or phone
Qwen3-VL-Embedding Alibaba Qwen 2.13B – 8.14B 32K Apache 2.0 (Permissive) 4.9 GB Laptop or phone
EmbeddingGemma 2 Google DeepMind 270M – 740M 8K Apache 2.0 (Permissive) 0.621 GB Laptop or phone
Nemotron 3.5 Embed NVIDIA 1.55B – 8.38B 32K OpenMDW-1.1 (Permissive) 1.46 GB Laptop or phone