# Qwen3-VL-Embedding
Qwen3-VL-Embedding is Qwen's embedding model built on Qwen3-VL. It maps images and text into one vector space, so a text query can find the page or photo it describes. LibreYOLO runs the 2B size on the embed task, the first VLM-based embedder in the library.
Tasks: Embeddings. Install: pip install "libreyolo[vlm]".
Verified against LibreYOLO v1.8.0.

## Install

Qwen3-VL-Embedding belongs to LibreYOLO's VLM tier and needs the `vlm` extra.

```bash
pip install "libreyolo[vlm]"
```

## Predict

Weights download on first use from `LibreYOLO/LibreQwen3VLEmbedding2b`, a
mirror of `Qwen/Qwen3-VL-Embedding-2B` with unchanged parameters. The
architecture runs through the transformers Qwen3-VL implementation.

**Python**

```python
from libreyolo import LibreVLM, SAMPLE_IMAGE

model = LibreVLM("qwen3-vl-embedding-2b")   # weights download on first use
result = model.predict(SAMPLE_IMAGE)
print(result.embeddings.shape)              # (1, 2048), unit length
```

**Retrieval**

```python
from libreyolo import LibreVLM

model = LibreVLM("qwen3-vl-embedding-2b")
pages = model.embed("pages/")               # (N, 2048), one row per image

# Queries carry a task instruction; images keep the default one.
q = model.embed_text(
    ["total due on the invoice"],
    prompt="Retrieve images or text relevant to the user's query.",
)                                           # (1, 2048)
scores = q @ pages.T                        # cosine similarity
```

The family loads through `LibreVLM()`, or the `LibreQwen3VLEmbedding` class.
It follows the same [embed](/docs/tasks/embeddings) contract as CLIP, SigLIP2 and
PE: `predict(image)` returns `Results.embeddings` of shape `(1, 2048)`,
`embed(source)` stacks one row per image, and `embed_text(texts, prompt=None)`
returns text rows in the same space. Rows are L2-normalized, so a dot product
is a cosine similarity.

Each input is a chat: a system instruction, then the image or the text. The
hidden state at the last token, normalized, is the vector. `prompt` is that
instruction. The default is upstream's `"Represent the user's input."`; set
`prompt=` at construction to change it for images and texts, or pass it to
`embed_text` per call. For retrieval, give queries a task instruction and keep
the default for documents.

Images are resized on a 32-pixel grid to between 4,096 and 1,843,200 pixels,
matching upstream's reference embedder. Against that embedder, in float32 on
CPU, text vectors differ by at most 7.4e-7 and image vectors have a cosine of
at least 0.99998.

## Variants

One size, `2b`. The 8B size is not wrapped. Not supported: combined
image-plus-text inputs, video clips (a video is embedded frame by frame), and
Matryoshka truncation as an option; slice the first `k` columns and
re-normalize instead.

## Export

Inference only: `chat()`, `set_classes()`, `train()`, `val()` and `export()`
raise.

## Licensing

Check the license on the Hugging Face repository of the specific weights you download. That repository is authoritative and licenses are not always uniform across a family. This is a description of the licenses involved, not legal advice.

- Original work: Qwen3-VL-Embedding, Alibaba (Qwen Team)
- Upstream license: Apache-2.0
- Upstream source: https://github.com/QwenLM/Qwen3-VL-Embedding
- LibreYOLO code: Apache-2.0
- Weights: Apache-2.0, republished at https://huggingface.co/LibreYOLO
- Interpretation: Apache-2.0 is a permissive license, so the Qwen3-VL-Embedding-2B weights can be used in commercial and closed-source products. Redistributing them means keeping the license text and attribution notices, and the license grants a patent license. LibreYOLO copies no model source: the Qwen3-VL architecture runs through the Apache-2.0 Hugging Face Transformers implementation, and libreyolo/models/vlm/qwen3vl_embedding.py (MIT) adapts only the input conventions and last-token pooling of the reference embedder in QwenLM/Qwen3-VL-Embedding at commit 393e2978d27852b0d0230d6994f37f9c15bed73c (Apache-2.0, see libreyolo/models/vlm/NOTICE). The weights are a mirrored snapshot of Qwen/Qwen3-VL-Embedding-2B at revision 9f2f7e710d6d81056aa5c0a4f04764fec6bb7bda, hosted at LibreYOLO/LibreQwen3VLEmbedding2b with learned parameters unchanged and not gated; the upstream Hugging Face repository ships no LICENSE file, so the mirror carries the Apache-2.0 text from the GitHub repository.

Qwen3-VL-Embedding-2B is Apache-2.0, by the
[Qwen team at Alibaba](https://github.com/QwenLM/Qwen3-VL-Embedding).
Commercial use is allowed. The mirror is not gated; because the upstream
Hugging Face repository ships no LICENSE file, it carries the Apache-2.0 text
from the GitHub repository.

## Citation

```bibtex
@article{qwen3vlembedding,
  title={Qwen3-VL-Embedding and Qwen3-VL-Reranker: A Unified Framework for State-of-the-Art Multimodal Retrieval and Ranking},
  author={Li, Mingxin and Zhang, Yanzhao and Long, Dingkun and Chen, Keqin and Song, Sibo and Bai, Shuai and Yang, Zhibo and Xie, Pengjun and Yang, An and Liu, Dayiheng and Zhou, Jingren and Lin, Junyang},
  journal={arXiv},
  year={2026}
}
```

Copied from https://github.com/QwenLM/Qwen3-VL-Embedding#citation
