Qwen3-VL-Embedding-2B vs Qwen3-VL-Embedding-8B

Qwen3-VL-Embedding-2B

A small multimodal embedding cousin that puts text, images, screenshots and video into one vector space, in 4.3 GB of BF16 files that fit any 8 GB machine, under plain Apache 2.0. It loads with Sentence Transformers, vLLM or SGLang. The main gaps are no quantised build from Alibaba, few publisher replies and no safety material.

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.

Qwen3-VL-Embedding-2B

1. Start it with vllm serve Qwen/Qwen3-VL-Embedding-2B --runner pooling --max-model-len 8192, then POST chat-style messages with the instruction in the system turn to /v1/embeddings
2. Or pip install -U sentence-transformers (5.4 or later) and SentenceTransformer('Qwen/Qwen3-VL-Embedding-2B').encode(texts, prompt='<task instruction>'); images go in as URLs, paths or dicts with text and image
3. Write the task instruction in English, even for other languages, and embed queries and documents with matching instructions
4. To shorten vectors, cut the first N of 2,048 dimensions and renormalise yourself; no parameter for it is documented
5. Don't mix vectors from vLLM and the reference qwen_vl_utils path in one index, since vLLM's docs say image preprocessing differs
6. 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