Qwen3-VL-Reranker

Alibaba's Apache 2.0 multimodal rerankers (2B and 8B) that score text, images, screenshots and video against a query.

In short. Qwen3-VL-Reranker re-sorts search results that may be text, pictures, screenshots or video. It reads a query and a candidate together and gives a relevance score. It comes in 2B and 8B sizes under Apache 2.0 and is meant to follow the Qwen3-VL-Embedding models.

Summary

Qwen3-VL-Reranker is the reranking half of Alibaba's Qwen3-VL-Embedding release of January 2026, built on the Qwen3-VL vision-language models and described in a technical report dated 8 January 2026 (arXiv 2601.04720). It takes a query and a document as input, either of which can be text, an image, a screenshot, a video or a mix such as text plus an image, and outputs a relevance score. There are 2B (28 layers) and 8B (36 layers) sizes, both with a 32K-token context, 30+ languages and custom instructions, and the card says instructions typically add 1% to 5%. Qwen intends the embedding model to do the first-stage recall and the reranker to refine the results.

Qwen's table gives the 2B and 8B rerankers MMEB-v2 retrieval averages of 75.1 and 79.2, video scores of 52.1 and 55.8, visual-document scores of 83.4 and 86.3, MMTEB retrieval scores of 70.0 and 74.9, JinaVDR scores of 80.9 and 83.6 and ViDoRe v3 scores of 60.8 and 66.7. The 2B reranker is slightly below the Qwen3-VL-Embedding-2B it re-scores on image (73.8 against 74.8) and video (52.1 against 53.6), and the jina-reranker-m0 baseline is ahead of it on visual documents (85.2 against 83.4 on VisDoc and 82.2 against 80.9 on JinaVDR). These are Qwen's own figures.

The license is Apache 2.0, and the BF16 weights are 4.26 GB for the 2B and 17.53 GB for the 8B. The card shows Sentence Transformers' CrossEncoder and a Transformers script that needs Transformers 4.57 or later, qwen-vl-utils 0.0.14 or later and PyTorch 2.8. In the 30 days to 11 November 2026, Hugging Face counted about 74,000 downloads of the 2B and 12,000 of the 8B.

Good for

  • Reranking results from mixed collections of text, page images, screenshots and video, after Qwen3-VL-Embedding or any other first stage
  • Visual document search, where the 8B scores 86.3 on VisDoc in Qwen's table
  • A 2B model for cheaper multimodal reranking
  • Commercial products, thanks to the Apache 2.0 license

Watch out for

  • The scores are Qwen's own, and in its table the 2B trails jina-reranker-m0 on two visual-document tests
  • Each query and candidate pair is scored separately, so image and video candidates make reranking costly
  • The card pins PyTorch 2.8.0 for its Transformers script, so use a separate environment
  • For text only, the lighter Qwen3-Reranker is the sibling to try first

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-Reranker. 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
Reranker-2B BF16 2.13B 4.26 GB 4.9 GB Laptop or phone
Reranker-8B BF16 8.77B 18 GB 20 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-Reranker?
Qwen3-VL-Reranker re-sorts search results that may be text, pictures, screenshots or video. It reads a query and a candidate together and gives a relevance score. It comes in 2B and 8B sizes under Apache 2.0 and is meant to follow the Qwen3-VL-Embedding models.
Can I use Qwen3-VL-Reranker commercially?
Yes. Qwen3-VL-Reranker 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-Reranker open source?
Qwen3-VL-Reranker 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-Reranker on my own computer?
Yes. The smallest variant, Reranker-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-Reranker released?
Qwen3-VL-Reranker was released by Alibaba Qwen on January 8, 2026.
How long a context window does Qwen3-VL-Reranker have?
Qwen3-VL-Reranker 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-VL-Embedding Alibaba Qwen 2.13B – 8.14B 32K Apache 2.0 (Permissive) 4.9 GB Laptop or phone
Qwen3-Reranker Alibaba Qwen 596M – 8.19B 32K Apache 2.0 (Permissive) 1.37 GB Laptop or phone
Nemotron 3.5 Rerank 8B NVIDIA 8.38B 32K OpenMDW-1.1 (Permissive) 19 GB Gaming GPU
Jina Reranker v3.5 Jina AI 597M 128K CC BY-NC (Non-commercial) 0.46 GB Laptop or phone