Qwen3-VL-Embedding logo

Qwen3-VL-Embedding

Visit

Qwen3-VL-Embedding is Qwen's Apache-2.0 multimodal embedding family (2B, 8B) for text, images and video, with 32K context and up to 4096 dimensions.

Share:
View alternatives

Qwen3-VL-Embedding is a family of multimodal embedding models from the Qwen team, in 2B and 8B sizes. According to the 8B model card, it is built on Qwen3-VL-8B-Instruct and embeds text, images, screenshots, videos and mixed inputs such as text plus image or text plus video into one vector space. The accompanying paper, arXiv 2601.04720, was published on January 8, 2026. The weights are released under the Apache-2.0 license.

Key features

  • Mixed-modality input: text, images, screenshots, video and combinations of them.
  • Long context: 32K tokens.
  • Flexible dimensions: up to 4096 dimensions for the 8B model and 2048 for the 2B model, with custom output sizes from 64 to 4096 supported through Matryoshka representation learning (MRL).
  • Instruction-aware: the model accepts task instructions with the input, and supports quantization of the output embedding.
  • Multilingual: the card states support for 30+ languages.
  • Reranker companion: Qwen3-VL-Reranker provides relevance scoring for the second stage of retrieval.

Model specs

Spec 2B 8B
Max embedding dimension 2048 4096
Context length 32K tokens 32K tokens
License Apache-2.0 Apache-2.0
Hugging Face repo Qwen/Qwen3-VL-Embedding-2B Qwen/Qwen3-VL-Embedding-8B

Vendor-reported results

On the MMEB-V2 benchmark, the 8B model card reports an overall score of 77.9, with 80.1 on image tasks, 66.1 on video tasks and 83.3 on visual document tasks. These are results published by the Qwen team.

Use cases

  • Searching screenshots, slides and scanned pages by text query.
  • Video and image retrieval for multimedia libraries.
  • Multimodal RAG where documents mix text and visuals.

Getting started

  1. Pick the 2B or 8B model from the model card.
  2. Use Sentence Transformers, Transformers (transformers 4.57.0 or later, qwen-vl-utils 0.0.14 or later, torch 2.8.0), vLLM or SGLang.
  3. Truncate to a smaller MRL dimension if index size matters, and test recall on your own data.
  4. Add a reranker for higher precision.

FAQ

What sizes does Qwen3-VL-Embedding come in?

2B and 8B.

What is the maximum embedding dimension?

4096 for the 8B model and 2048 for the 2B model, with custom sizes from 64 to 4096 supported.

What license is it under?

Apache-2.0.

Does it handle video?

Yes. It embeds text, images, screenshots, videos and mixed inputs.

Alternatives

Conclusion

Qwen3-VL-Embedding brings text, image and video retrieval into one open, Apache-2.0 model family with flexible vector sizes. Read the official model card for usage, and browse the models category for related models.

Comments

No comments yet. Be the first to comment!