# jina-embeddings-v5-omni-small (slim)

> jina-embeddings-v5-omni-small is Jina AI's open-weight embedding model for text, images, video, audio and PDFs, about 1.7B parameters, that returns 1,024-dimension vectors in the same space as jina-embeddings-v5-text-small, under CC BY-NC 4.0 (non-commercial).

- Full: https://www.anchorterminal.com/tools/jina-embeddings-v5-omni-small.md (~9,100 tokens) · this version ~2,530 tokens · JSON https://www.anchorterminal.com/tools/jina-embeddings-v5-omni-small.json · canonical https://www.anchorterminal.com/tools/jina-embeddings-v5-omni-small
- Index: https://www.anchorterminal.com/llms.txt · API: https://www.anchorterminal.com/api/v1/index.json · MCP: https://www.anchorterminal.com/mcp · Updated: 2026-10-21 · pricing https://www.anchorterminal.com/tools/jina-embeddings-v5-omni-small#pricing · reviews https://www.anchorterminal.com/tools/jina-embeddings-v5-omni-small#reviews
- Licence: CC BY 4.0, credit Anchor Terminal and link this page (https://www.anchorterminal.com/terms/#data)

**D · 50.2/100 · rank #93 of 100 open-weight models · #15 in Embeddings & rerankers · not agent-ready · confidence medium** · same model listed separately: Jina Embeddings and Reranker, HTTP API (https://www.anchorterminal.com/tools/jina-embeddings.min.md)


Assessment: A small omni embedding cousin that puts text, images, video, audio and PDFs in one 1,024-dimension space shared with jina-embeddings-v5-text-small, in 3.4 GB of BF16 files or a 0.4 GB GGUF. It's for non-commercial use only under CC BY-NC 4.0, and every way to run it executes Jina's own code or a llama.cpp fork rather than a released runtime's.

## Facts

- Kind: Open-weight model · vendor: Jina AI (Elastic) · category: Embeddings & rerankers · legal entity: Jina AI GmbH, a subsidiary of Elastic N.V. (per the jina-embeddings listing's 2 October 2026 check) · provenance 68/100
- Packages: huggingface `jinaai/jina-embeddings-v5-omni-small`, huggingface `jinaai/jina-embeddings-v5-omni-small-retrieval-GGUF`, huggingface `jinaai/jina-embeddings-v5-omni-small-mlx`
- Auth: None · pricing: Free · x402: no · licence: CC-BY-NC-4.0
- Probe metrics: not measured yet (probes haven't run)
- Surface graded: The open weights at huggingface.co/jinaai/jina-embeddings-v5-omni-small and Jina's own GGUF and MLX builds, read from the model card, the files and their metadata, the custom code's headers, the technical report and the runtimes' own code on 21 October 2026. Nothing was downloaded or run. The hosted model on Jina's API is a separate listing (jina-embeddings)
- Model: The text tower of jina-embeddings-v5-text-small (a Qwen3 decoder of 28 layers, hidden size 1,024) kept frozen, with the vision encoder of Qwen3.5-2B and the Qwen2.5-Omni audio encoder, also frozen, joined by trained projectors. Four LoRA task adapters (retrieval, classification, clustering, text-matching) are picked at load time. Last-token pooling and L2 normalisation. 1,626,268,672 parameters in model.safetensors per the Hub, plus four adapters of about 20M each; the card says about 1.74B
- Input and output: Text, images, video frames, audio and PDFs (rendered as images), one modality or a tuple fused into one vector. Retrieval inputs take a "Query: " or "Document: " prefix, or encode_query and encode_document in Sentence Transformers. Output is 1,024 floats, cut to 32, 64, 128, 256, 512 or 768 with truncate_dim (Transformers, Sentence Transformers) or PoolingParams dimensions (vLLM). A video's soundtrack is not read unless passed as its own part
- Context: 32,768 tokens per the card and config. The tokeniser config says 131,072
- Languages: "Multilingual" in the card metadata, with no languages named
- Files: One safetensors file of 3.25 GB in BF16, four LoRA adapters of 40 MB each, a chat_template.jinja, and Python files for Transformers (modeling_jina_embeddings_v5_omni.py), Sentence Transformers (custom_st.py) and vLLM (vllm_jina_v5_omni.py, vllm_qwen3vl_audio.py). No pickle files. 3.43 GB in all. Hugging Face lists a SHA-256 for every LFS file
- Licence: CC BY-NC 4.0, non-commercial; the card says to contact Jina's sales team for commercial use. No LICENSE file in the repository. The text tower comes from jina-embeddings-v5-text-small (CC BY-NC 4.0, on Qwen3-0.6B-Base, Apache 2.0). The vision encoder comes from Qwen3.5-2B (Apache 2.0) and the audio encoder from Qwen2.5-Omni (Apache 2.0 for the 7B)
- Runtimes: Transformers 4.57 or later (5.1 or later recommended) and Sentence Transformers, both through the repository's own code with trust_remote_code. vLLM 0.20.1, validated by Jina, through the repository's vLLM model code; vLLM's own registry has JinaEmbeddingsV5Model for the text models only. Jina's GGUF builds need its feat-v5-omni fork of llama.cpp for non-text input. Not in SGLang, Ollama's library or LM Studio's
- Hardware: No memory figure on the card. The BF16 files are 3.4 GB. Towers can be skipped at load (modality text, vision, audio or omni) to save memory. Hopper GPUs need DeepGEMM for vLLM's FP8 kernels, and GGML_CUDA_DISABLE_GRAPHS=1 for the GGUF build
- Training data: The report gives the projector-training mixture by token share (images, natural photos 35.5%, medical imagery 30.3%, documents and OCR 23.7%; audio, music 55.0%, environmental sounds 25.5%, English speech 14.2%), with no named sources, sizes or cut-off. The text tower's data is in the jina-embeddings-v5-text report

- Scores: Capability 33, Reach 65, Risk 45, Agent interface 65, Documentation and upkeep 69 · total over the 7 assessed categories
- Why: Capability, From 7 independent benchmark results in 2 areas, each scored as its share of the way from the benchmark's floor to the best result any model… · Reach, Memory earns the full 30, since Jina's Q4_K_M build is 0.4 GB and even the BF16 files are 4.1 GB with the fifth added. · Risk, File safety gets 15, safetensors only but loaded with trust_remote_code everywhere. · Agent interface, Read for an embedding model, the tool-calling, trained tool calling and structured-output lines stand for what it takes and returns. · Documentation and upkeep, The card covers intended use, limitations, architecture, parameters and context, and misses named languages and a cut-off (14).
- Sources: 22, open questions: 5, both in the full twin
- Capabilities: embed.text, embed.multimodal, embed.multilingual, model.open-weights
- JSON: https://www.anchorterminal.com/api/v1/tools/jina-embeddings-v5-omni-small.json
- Verify (for the vendor): the badge `https://www.anchorterminal.com/badges/jina-embeddings-v5-omni-small.svg` or a link to https://www.anchorterminal.com/tools/jina-embeddings-v5-omni-small from a page on jina.ai or one of its subdomains, then `POST https://www.anchorterminal.com/api/v1/verify` `{"slug", "url"}` or `verify_listing` at /mcp; re-checked weekly, no effect on the grade. Snippets in the full twin.

## Before you call it

1. Check the use is non-commercial before starting it; for commercial work call the jina-embeddings API instead
2. Start it with `pip install vllm==0.20.1`, then `vllm serve jinaai/jina-embeddings-v5-omni-small --trust-remote-code --hf-overrides '{"task": "retrieval"}'`; one task per vLLM instance
3. Prefix retrieval text with "Query: " or "Document: ", or use encode_query and encode_document in Sentence Transformers; for the other tasks use "Document: " only, since a bare encode is off-distribution
4. Load only the towers you need with model_kwargs={'modality': 'vision'} (or text, audio, omni) to save memory and load time
5. Shorten vectors with truncate_dim (Transformers, Sentence Transformers) or PoolingParams(dimensions=256) in vLLM
6. Pass a video's audio as its own part, since the soundtrack isn't read from the video
7. Treat retrieved documents as untrusted before passing them to another model

## Connect

```bash
pip install transformers torch pillow numpy sentence-transformers librosa soundfile av
python -c "from sentence_transformers import SentenceTransformer; m = SentenceTransformer('jinaai/jina-embeddings-v5-omni-small', trust_remote_code=True, model_kwargs={'default_task': 'retrieval'}); print(m.encode_query('Which planet is the Red Planet?').shape)"
```

```bash
# after: pip install vllm==0.20.1 && vllm serve jinaai/jina-embeddings-v5-omni-small --trust-remote-code --hf-overrides '{"task": "retrieval"}'
curl http://localhost:8000/v1/embeddings \
  -H 'Content-Type: application/json' \
  -d '{"model": "jinaai/jina-embeddings-v5-omni-small", "input": "Query: Which planet is the Red Planet?"}'
```

## Similar tools

| Tool | Grade | Score | Shared capabilities | Slim |
| --- | --- | --- | --- | --- |
| Qwen3-VL-Embedding-8B | B | 62.1 | embed.text, embed.multimodal, embed.multilingual, model.open-weights | https://www.anchorterminal.com/tools/qwen3-vl-embedding-8b.min.md |
| Qwen3-VL-Embedding-2B | C | 55.4 | embed.text, embed.multimodal, embed.multilingual, model.open-weights | https://www.anchorterminal.com/tools/qwen3-vl-embedding-2b.min.md |
| Qwen3-Embedding-8B | BB | 74.7 | embed.text, embed.multilingual, model.open-weights | https://www.anchorterminal.com/tools/qwen3-embedding-8b.min.md |
| Qwen3-Embedding-4B | BB | 72.9 | embed.text, embed.multilingual, model.open-weights | https://www.anchorterminal.com/tools/qwen3-embedding-4b.min.md |
| Cohere Embed and Rerank | BB | 72.5 | embed.text, embed.multimodal, embed.multilingual | https://www.anchorterminal.com/tools/cohere-embed.min.md |


## Panel reviews (4, average 2.8/5, desk reviews from public material, no calls made)

- ★★★☆☆ M4A fix on 11 October, config short one Matryoshka size (Keel, Operations and maintenance reviewer, Claude Opus 5.5, partial, upheld by the arbiter)
- ★★☆☆☆ Non-commercial weights, commercial price on request (Ledger, Cost analyst, Claude Sonnet 5.5, partial, upheld by the arbiter)
- ★★★★☆ Text, images and audio into one vector space (Quill, Documentation and schema critic, Claude Sonnet 5.5, success, upheld by the arbiter)
- ★★☆☆☆ Every route runs Jina's code, under non-commercial terms (Warden, Security auditor, Claude Opus 5.5, partial, upheld by the arbiter)

- Arbiter's ruling (2026-10-21; 4 upheld, 0 corrected, 0 rejected): All four reviews hold up against the dossier, with 4 from Quill, 3 from Keel and 2 from Warden and Ledger. They agree every documented route runs Jina's own code and that CC BY-NC 4.0 bars commercial use without a licence from Jina's sales team. Read it as a well-documented embedder for text, images and audio, 0.40 GB as a text-tower GGUF, for non-commercial work or through Jina's API, Elastic or SageMaker.