Qwen3-VL-Reranker

Open weights · Qwen, Alibaba · Updated 15 Oct 2026

Qwen3-VL-Reranker (2B / 8B) scores a query against documents that can be text, images, screenshots, video, or a mix of them. A scanned page or a slide is ranked as the picture it is, not through whatever OCR managed to extract. It is built on Qwen3-VL, with a 32K context, 30+ languages and an Apache 2.0 licence. For text-only documents, its sibling Qwen3-Reranker is the one to look at.

Model variants

Both take a task instruction. Qwen report that instructions usually add 1–5%, and recommend writing them in English even for other languages. The family pairs with Qwen3-VL-Embedding, the first-stage retriever for the same kinds of input.

Scores, and whose they are

Qwen’s own runs, from the model card. The card doesn’t say which first stage each reranker reordered. MMEB-v2 covers image, video and visual-document retrieval; JinaVDR and ViDoRe v3 are visual-document benchmarks (page images); MMTEB retrieval is text. None of them is BEIR.

In Qwen’s table, the 2B is close to jina-reranker-m0, Jina’s multimodal reranker, whose weights are licensed for non-commercial use only: ahead on images and ViDoRe v3, behind on visual documents and JinaVDR. The 8B leads every column. One outside run: LightOn, who make a competing model, report BEIR over BM25’s top 100, averaged over 11 datasets kept out of their own training data. There the 2B scores 49.08 and jina-reranker-m0 51.97: on text, the 2B trails m0 in that run.

Quick start

sentence-transformers

pip install sentence-transformers

from sentence_transformers import CrossEncoder

model = CrossEncoder("Qwen/Qwen3-VL-Reranker-2B")

query = "A woman playing with her dog on a beach at sunset."
documents = [
    "A woman shares a joyful moment with her golden retriever on a beach at sunset.",
    "https://qianwen-res.oss-cn-beijing.aliyuncs.com/Qwen-VL/assets/demo.jpeg",
    {"text": "Golden retriever on the beach", "image": "https://qianwen-res.oss-cn-beijing.aliyuncs.com/Qwen-VL/assets/demo.jpeg"},
]

prompt = "Retrieve images or text relevant to the user's query."
rankings = model.rank(query, documents, prompt=prompt)

A document can be text, an image URL or path, or a dict with both. Without prompt, the instruction is “Retrieve text relevant to the user’s query.” The model card also shows a plain transformers path, which needs a helper script from Qwen’s GitHub repository, and a vLLM path.

Hosted options

Alibaba Cloud Model Studio sells qwen3-vl-rerank, and SiliconFlow hosts Qwen3-VL-Reranker-8B. Both are reachable from mainland China; see reranking in mainland China for endpoints and prices. Alibaba Cloud publishes no scores for its hosted version.

When to pick Qwen3-VL-Reranker

• Yes: your documents are page images: scans, slide decks, screenshots, dashboards.
• Yes: a query should match images or video, or documents that mix text and pictures.
• Maybe not: the documents are plain text. Use Qwen3-Reranker or another reranker built for text; it does the same job with less compute.
• Maybe not: CPU-only serving. Even the 2B wants a GPU, and images make every pair more expensive to score than text.

Pros and cons

Other models