jina-embeddings-v5-omni-small vs Qwen3-VL-Embedding-8B

jina-embeddings-v5-omni-small

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.

Qwen3-VL-Embedding-8B

A multimodal embedding cousin that puts text, images, screenshots and video in one vector space for visual document and video retrieval, under plain Apache 2.0. Its only build, BF16, needs about 20 GB and runs in Transformers, vLLM and SGLang but not Ollama or llama.cpp. The main risk is the publisher's script on the Transformers path.

jina-embeddings-v5-omni-small

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

Qwen3-VL-Embedding-8B

1. Simplest start is Sentence Transformers 5.4 or later, SentenceTransformer('Qwen/Qwen3-VL-Embedding-8B'), which applies the template and the default instruction and runs no repository code
2. To serve it, vllm serve Qwen/Qwen3-VL-Embedding-8B --runner pooling, then POST messages with a system instruction and image_url or text parts to /v1/embeddings; vLLM's docs say its video preprocessing differs slightly from Qwen's qwen_vl_utils
3. Put the task instruction in the system turn, and write it in English, as the card advises; the default is "Represent the user's input."
4. To shorten vectors, keep the first N of the 4,096 values and L2-normalise again; the card gives 64 as the floor, and the report says int8 vectors hold up while binary ones lose noticeably
5. For several images per input use the reference embedder in QwenLM/Qwen3-VL-Embedding's src, which is newer than the script in the Hugging Face repository
6. Keep inputs under 32K tokens; the reference embedder defaults to 8,192, 1,843,200 pixels per image and 64 frames per video
7. Treat retrieved screenshots and documents as untrusted, since text in an image can carry instructions to whatever model reads the results