Qwen3-VL-Embedding

Alibaba's Apache 2.0 multimodal embedding models (2B and 8B) that put text, images, screenshots and video in one search space.

In short. Qwen3-VL-Embedding turns text, pictures, screenshots and video into numbers in one shared space, so you can search a pile of slides or videos with a sentence, or find text with a picture. It comes in 2B and 8B sizes under Apache 2.0, and a matching reranker series re-scores the best results.

Summary

In January 2026 Alibaba's Qwen team released Qwen3-VL-Embedding and Qwen3-VL-Reranker, built on its Qwen3-VL vision-language models and described in a technical report dated 8 January 2026 (arXiv 2601.04720). Both come in 2B and 8B sizes under the Apache 2.0 license. The embedding models accept text, images, screenshots, video and mixtures such as text plus an image, in 30+ languages with a 32K-token context. The 2B model outputs 2,048-dimensional vectors and the 8B model up to 4,096, and the card says the 8B model supports any size from 64 to 4,096 dimensions.

The two parts work as a pair. The embedding model does the fast first-stage recall, and the reranker takes a query and a document, either of which can be text, an image, video or a mix, and returns a relevance score. Both are instruction-aware: Qwen says a task-specific instruction typically improves results by 1% to 5%, and it recommends writing instructions in English even for other languages. The embedding output can also be quantized to save storage.

On the MMEB-V2 multimodal benchmark, Qwen's card reports an overall 77.9 for the 8B model and 73.4 for the 2B model, and the report says the 8B model ranked first among the models it compared as of 8 January 2026. These are developer-reported, and leaderboards move quickly. The BF16 weights are about 16.3 GB for 8B and 4.3 GB for 2B. If you only search text, the text-only Qwen3-Embedding family is cheaper to run and covers more languages.

Good for

  • Searching images, screenshots, slides and other document images with text queries
  • Cross-modal retrieval such as text to image, text to video or image to text
  • Two-stage multimodal search with the matching Qwen3-VL-Reranker models
  • Choosing the vector size to trade accuracy for storage (8B supports 64 to 4,096 dimensions)

Watch out for

  • Heavier than text-only embedding models, with only 2B and 8B sizes
  • Covers 30+ languages, fewer than the 100+ of the text-only Qwen3-Embedding
  • The benchmark ranking is Qwen-reported and dates from January 2026
  • Vectors are not interchangeable between models, so switching means re-indexing

Performance claims come from the developer and have not been independently verified.

How much memory does it need?

Memory figures for each variant of Qwen3-VL-Embedding. Image, video, speech and music pipelines also load a text encoder and a decoder, so the memory needed can be larger than the main checkpoint.

Variants Parameters Download size Memory needed Fits on
Embedding-2B BF16 2.13B 4.26 GB 4.9 GB Laptop or phone
Embedding-8B BF16 8.14B 16 GB 19 GB Gaming GPU

Download size covers every component needed to run the model, such as the generator, text encoder and decoder. Memory needed is a stated or published measured minimum where one exists; otherwise it is the download size plus about 15%. Larger images, longer clips and higher resolutions need more.

The license in plain words

One of the most widely used permissive licenses. You may use, change and sell products built on the model with only light conditions.

More about Apache 2.0

See also

Frequently asked questions

What is Qwen3-VL-Embedding?
Qwen3-VL-Embedding turns text, pictures, screenshots and video into numbers in one shared space, so you can search a pile of slides or videos with a sentence, or find text with a picture. It comes in 2B and 8B sizes under Apache 2.0, and a matching reranker series re-scores the best results.
Can I use Qwen3-VL-Embedding commercially?
Yes. Qwen3-VL-Embedding is released under Apache 2.0, a permissive license. You can use it commercially if you follow the standard conditions, such as keeping the license notice.
Is Qwen3-VL-Embedding open source?
Qwen3-VL-Embedding is released under Apache 2.0, a license approved by the Open Source Initiative. Training data and training code may not be public, so it is best described as open weights under an open-source license.
Can I run Qwen3-VL-Embedding on my own computer?
Yes. The smallest variant, Embedding-2B, needs about 4.9 GB of memory at 4-bit precision, which fits a modern laptop or a mid-range graphics card.
When was Qwen3-VL-Embedding released?
Qwen3-VL-Embedding was released by Alibaba Qwen on January 2026.
How long a context window does Qwen3-VL-Embedding have?
Qwen3-VL-Embedding supports a context window of up to 32K tokens.

Official links

Sources

Specifications come from these primary sources. Benchmark claims, where mentioned, are reported by the developer.

Release date and facts last verified November 11, 2026.

Related models

Model Released Parameters Context License Smallest variant needs
Qwen3-Embedding Alibaba Qwen 600M – 7.57B 32K Apache 2.0 (Permissive) 1.38 GB Laptop or phone
EmbeddingGemma 2 Google DeepMind 270M – 740M 8K Apache 2.0 (Permissive) 0.621 GB Laptop or phone
Jina Embeddings v5 Jina AI 239M – 677M 32K CC BY-NC (Non-commercial) 0.552 GB Laptop or phone
BGE-M3 BAAI 568M 8K MIT (Permissive) 2.61 GB Laptop or phone