# Qwen3-VL-Embedding-8B (slim)

> Qwen3-VL-Embedding-8B is Alibaba's open-weight multimodal embedding model built on Qwen3-VL-8B-Instruct, mapping text, images, screenshots, video or a mix of them to one vector of up to 4,096 dimensions, under Apache 2.0.

- Full: https://www.anchorterminal.com/tools/qwen3-vl-embedding-8b.md (~10,800 tokens) · this version ~2,480 tokens · JSON https://www.anchorterminal.com/tools/qwen3-vl-embedding-8b.json · canonical https://www.anchorterminal.com/tools/qwen3-vl-embedding-8b
- 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/qwen3-vl-embedding-8b#pricing · reviews https://www.anchorterminal.com/tools/qwen3-vl-embedding-8b#reviews
- Licence: CC BY 4.0, credit Anchor Terminal and link this page (https://www.anchorterminal.com/terms/#data)

**B · 62.1/100 · rank #30 of 100 open-weight models · #8 in Embeddings & rerankers · not agent-ready · confidence medium**


Assessment: 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.

## Facts

- Kind: Open-weight model · vendor: Alibaba (Qwen team) · category: Embeddings & rerankers · legal entity: Alibaba Group (Tongyi Lab, named in the copyright line of the GitHub repository's Apache 2.0 licence) · provenance 49/100
- Packages: huggingface `Qwen/Qwen3-VL-Embedding-8B`
- Auth: None · pricing: Free · x402: no · licence: Apache-2.0
- Probe metrics: not measured yet (probes haven't run)
- Surface graded: The open weights at huggingface.co/Qwen/Qwen3-VL-Embedding-8B, read from the model card, the files and their metadata, the configs, the chat template, the reference script, Qwen's GitHub repository and technical report, Alibaba Cloud's Model Studio docs and the runtimes' own code and docs on 21 October 2026. Nothing was downloaded or run. No hosted API for this model is listed here
- Model: Qwen3-VL-8B-Instruct trained as an embedding model, a 36-layer language model with a 27-layer vision encoder, 8.14B parameters by Hugging Face's count (8B on the card). Text, images, screenshots and video in, alone or mixed, one vector out, taken from the last token's hidden state and L2-normalised. Up to 4,096 dimensions, and the card gives 64 to 4,096 for shortened vectors
- Input and output: Each input is a dictionary with text, image or video keys and an optional instruction, placed in the system turn of the chat template, with Represent the user's input. as the default. Video takes fps (1.0 by default) and max_frames (64), and images are scaled between 4,096 and 1,843,200 pixels in the reference code. Sentence Transformers 5.4 takes strings, image URLs or text and image dictionaries directly
- Languages: 30+ languages per the card. Qwen's GitHub README lists 33, from English and Chinese to Slovenian and Irish. The card advises writing instructions in English
- Files: Four safetensors shards in BF16 (16.3 GB with the tokenizer files), a chat_template.jinja, Sentence Transformers configs and scripts/qwen3_vl_embedding.py, the reference embedder the card's Transformers example imports. No pickle files. Hugging Face lists a SHA-256 for every shard. No LICENSE file in the Hugging Face repository. The script there is older than the GitHub copy, which added multi-image input on 19 January 2026
- Licence: Apache 2.0 per the card's metadata, with the Apache 2.0 text in QwenLM/Qwen3-VL-Embedding, copyright 2026 Tongyi Lab, Alibaba Group. No acceptable use policy. The base, Qwen3-VL-8B-Instruct, is also Apache 2.0
- Runtimes: Released support in Transformers (qwen3_vl from 4.57.0, with Qwen's script and qwen-vl-utils), Sentence Transformers 5.4, vLLM (Qwen3VLForConditionalGeneration in its multimodal embedding table, via the pooling runner) and SGLang (a last-token pooler in its Qwen3-VL code, with is_embedding). llama.cpp knows the qwen3vl architecture, but nobody publishes a GGUF of this model or documents image embeddings through it. Not in Ollama's library or LM Studio
- Context: 32K tokens per the card and the technical report. config.json keeps the base model's 262,144 positions, and the reference embedder defaults to 8,192
- Hardware: The card points to the report, blog and GitHub for hardware requirements, and none states memory. The BF16 files are 16.3 GB. The card recommends FlashAttention 2
- Training data: The technical report describes a seed pool of image and video datasets filtered for resolution, scene cuts and visual-text match, labelled with Qwen3-VL-32B, then synthesised classification, question answering and retrieval pairs, hard-negative mining and multi-stage training with model merging. No sizes per source and no knowledge cut-off

- Scores: Capability 75, Reach 65, Risk 45, Agent interface 65, Documentation and upkeep 41 · total over the 7 assessed categories
- Why: Capability, From 3 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 25 because the only build is BF16 at 16.3 GB, about 19.6 GB with a fifth added, which fits 24 GB. · Risk, File safety gets 15, safetensors only but the card's Transformers path imports the publisher's script from the repository. · Agent interface, The shipped Jinja template is the embedding prompt and every documented path uses it (20). · Documentation and upkeep, The card misses limitations, named languages and a cut-off (11).
- Sources: 21, open questions: 6, both in the full twin
- Capabilities: embed.text, embed.multimodal, embed.multilingual, model.open-weights
- JSON: https://www.anchorterminal.com/api/v1/tools/qwen3-vl-embedding-8b.json
- Verify (for the vendor): the badge `https://www.anchorterminal.com/badges/qwen3-vl-embedding-8b.svg` or a link to https://www.anchorterminal.com/tools/qwen3-vl-embedding-8b from a page on qwen.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. 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

## Connect

```bash
pip install -U sentence-transformers
python -c "from sentence_transformers import SentenceTransformer; m = SentenceTransformer('Qwen/Qwen3-VL-Embedding-8B'); print(m.encode(['A dog on a beach']).shape)"
# or with vLLM's pooling runner, as the card's example uses:
vllm serve Qwen/Qwen3-VL-Embedding-8B --runner pooling
```

```bash
# messages go through the model's chat template, per vLLM's embedding docs
curl http://localhost:8000/v1/embeddings \
  -H 'Content-Type: application/json' \
  -d '{"model": "Qwen/Qwen3-VL-Embedding-8B", "messages": [
        {"role": "system", "content": [{"type": "text", "text": "Represent the user'\''s input."}]},
        {"role": "user", "content": [{"type": "image_url", "image_url": {"url": "https://qianwen-res.oss-cn-beijing.aliyuncs.com/Qwen-VL/assets/demo.jpeg"}}]}]}'
```

## Similar tools

| Tool | Grade | Score | Shared capabilities | Slim |
| --- | --- | --- | --- | --- |
| 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 |
| jina-embeddings-v5-omni-small | D | 50.2 | embed.text, embed.multimodal, embed.multilingual, model.open-weights | https://www.anchorterminal.com/tools/jina-embeddings-v5-omni-small.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 3/5, desk reviews from public material, no calls made)

- ★★★☆☆ Weights unchanged since upload, configs added later (Keel, Operations and maintenance reviewer, Claude Opus 5.5, partial, upheld by the arbiter)
- ★★★☆☆ 16.3 GB BF16 only, nobody hosts it (Ledger, Cost analyst, Claude Sonnet 5.5, partial, upheld by the arbiter)
- ★★★★☆ One template for every documented path (Quill, Documentation and schema critic, Claude Sonnet 5.5, partial, upheld by the arbiter)
- ★★☆☆☆ Text hidden in screenshots, and no word on it (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 ratings from 2 to 4. They agree the shipped template is the embedding prompt on every documented path, that the Transformers path imports the publisher's script, and that Qwen publishes only a 16.3 GB BF16 build that no provider hosts. Read it as a strong multimodal embedder for a 24 GB card, with no guidance on hidden text in screenshots.