Qwen3-Embedding-0.6B vs Qwen3-VL-Embedding-8B

Qwen3-Embedding-0.6B

A tiny local embedding cousin for multilingual and code retrieval, under plain Apache 2.0, that fits in under 1 GB at Q8_0 and runs in llama.cpp, Ollama, vLLM, SGLang and Transformers. Queries need a task instruction and last-token pooling. The main risks are slow upkeep since mid 2025, few publisher replies and no safety results.

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-Embedding-0.6B

1. Start it with ollama pull qwen3-embedding:0.6b, then POST to /api/embed with the text in input
2. With llama.cpp, run llama-server -hf Qwen/Qwen3-Embedding-0.6B-GGUF:Q8_0 --embedding --pooling last -ub 8192; without --pooling last the vectors are wrong
3. Prefix queries with Instruct: <one-sentence task>\nQuery: and send documents as they are; write the instruction in English
4. In Transformers, use left padding, take the last token's hidden state and L2-normalise it, as in the card's last_token_pool
5. In Sentence Transformers, pass prompt_name="query" for queries; the prompt is stored in config_sentence_transformers.json
6. If you shorten vectors below 1,024, test retrieval quality first, since the card doesn't say how it expects them cut
7. Retrieved documents are untrusted text; embedding them doesn't remove instructions an attacker put in them

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