Qwen3-VL-Reranker
Qwen3-VL-Reranker is a multimodal reranking model family from Alibaba's Qwen team, published on Hugging Face as Qwen3-VL-Reranker-2B and Qwen3-VL-Reranker-8B. The 8B card's citation is dated 2026, with the paper (arXiv:2601.04720) listed as published on January 8. The 2B model is fine-tuned from Qwen3-VL-2B-Instruct (28 layers) and the 8B from Qwen3-VL-8B-Instruct. Both support a 32K context, more than 30 languages and the Apache-2.0 license.
Status
Live. Both cards are on Hugging Face. The 8B card pairs the reranker with the Qwen3-VL-Embedding models: the embedding model handles first-stage recall, and the reranker refines the results in a second stage. See the catalog entry for Qwen3-VL-Embedding.
Key Features
- Mixed-modality inputs: text, images, screenshots, video and combinations such as text plus image, for both the query and the documents.
- Two sizes: 2B and 8B parameters (the 8B's safetensors metadata lists 9B).
- Long context: 32K tokens.
- Multilingual: 30+ languages.
- Vendor-reported scores: on retrieval benchmarks, the 8B card lists MMEB-v2 average 79.2 (image 80.7, video 55.8, VisDoc 86.3), MMTEB 74.9, JinaVDR 83.6 and ViDoRe v3 66.7. The 2B card lists MMEB-v2 average 75.1. The cards claim the reranker beats the base embedding model and baseline rerankers, with 8B best on most tasks. These are Qwen's own results.
- License: Apache 2.0.
Use Cases
- Visual document retrieval, such as slides, scanned pages and screenshots.
- Video and image search where text-only rerankers cannot see the content.
- Second-stage ranking after a multimodal embedding search.
Limitation: it needs recent library versions (the card lists transformers 4.57.0 or later, qwen-vl-utils 0.0.14 or later and torch 2.8.0 for the helper script).
Pricing
The weights are free under Apache 2.0, so the cost is your own compute. This page does not quote hosted prices.
Getting Started
- Load it with Sentence Transformers as a
CrossEncoderand usepredictorrank. - Or use the Transformers
Qwen3VLRerankerhelper script from the card, or vLLM with the pooling runner and the score API. - Retrieve candidates with Qwen3-VL-Embedding, then rerank the top results.
Frequently Asked Questions
What sizes does Qwen3-VL-Reranker come in?
2B and 8B, each with a 32K context.
What inputs can it rank?
Text, images, screenshots, video and mixed combinations.
Is it free for commercial use?
Both cards list Apache-2.0.
Which embedding model does it pair with?
The 8B card pairs it with Qwen3-VL-Embedding for first-stage recall.
Alternatives
- Qwen3-VL-Embedding: the matching embedding model for first-stage retrieval.
- Jina AI Reranker v3: a text-focused multilingual reranker with a long context.
- Cohere Rerank 3.5: Cohere's hosted reranker.
Conclusion
Qwen3-VL-Reranker brings Apache-licensed, vendor-benchmarked reranking to text, images and video. Start with the 8B card or the 2B card and validate on your own retrieval set.
Sources
- Websitehttps://huggingface.co/Qwen/Qwen3-VL-Reranker-2B
- Social Mediahttps://x.com/Alibaba_Qwen
Comments
No comments yet. Be the first to comment!