Qwen3-VL-Embedding and Qwen3-VL-Reranker Release
Qwen has announced the release of Qwen3-VL-Embedding and Qwen3-VL-Reranker, a new series of models specifically engineered for multimodal information retrieval and cross-modal understanding. Built on the Qwen3-VL foundation models, these tools enable the seamless handling of text, images, screenshots, and video within a unified framework to improve the accuracy of multimodal content retrieval.
Technical Architecture
Qwen3-VL-Embedding and Qwen3-VL-Reranker utilize different architectural approaches to balance efficiency and precision in a retrieval pipeline.
Qwen3-VL-Embedding (Dual-Tower)
The Embedding model uses a dual-tower architecture to map single-modal or mixed-modal inputs into high-dimensional semantic vectors. It extracts the hidden state vector of the [EOS] token from the base model's last layer to serve as the final semantic representation. This independent encoding allows for the efficient, large-scale retrieval of candidates from a vector space.
Qwen3-VL-Reranker (Single-Tower)
The Reranker uses a single-tower architecture with a Cross-Attention mechanism to perform joint encoding of a (Query, Document) pair. By analyzing the interaction between the query and the document, the model outputs a precise relevance score based on the probability of generating special tokens (yes and no).
Model Specifications
Both the Embedding and Reranker series are available in 2B and 8B parameter sizes. All models support a sequence length of 32K and are instruction-aware, allowing for task-specific customization of input instructions.
| Model | Size | Layers | Seq Length | Embedding Dim | Quantization | MRL Support |
|---|---|---|---|---|---|---|
| Qwen3-VL-Embedding-2B | 2B | 28 | 32K | 2048 | Yes | Yes |
| Qwen3-VL-Embedding-8B | 8B | 36 | 32K | 4096 | Yes | Yes |
| Qwen3-VL-Reranker-2B | 2B | 28 | 32K | - | - | - |
| Qwen3-VL-Reranker-8B | 8B | 36 | 32K | - | - | - |
Performance and Benchmarks
Qwen3-VL-Embedding-8B achieves state-of-the-art (SOTA) results on the MMEB-v2 benchmark, surpassing previous open-source and proprietary models. It consistently performs at the top across image, visual document, and video retrieval subtasks.
On the text-only multilingual MMTEB benchmark, the Qwen3-VL-Embedding model is highly competitive with other similarly sized models, although it shows a performance gap compared to the text-only Qwen3-Embedding model.
Reranker Evaluation
The Qwen3-VL-Reranker models consistently outperform the base embedding models and baseline rerankers across various datasets, including JinaVDR and ViDoRe v3. The 8B variant provides the highest performance across most tasks.
| Model | MMEB-v2 (Avg) | MMEB-v2 (Image) | MMEB-v2 (Video) | MMEB-v2 (VisDoc) | MMTEB | JinaVDR | ViDoRe (v3) |
|---|---|---|---|---|---|---|---|
| Qwen3-VL-Embedding-2B | 73.4 | 74.8 | 53.6 | 79.2 | 68.1 | 71.0 | 52.9 |
| jina-reranker-m0 | - | 68.2 | - | 85.2 | - | 82.2 | 57.8 |
| Qwen3-VL-Reranker-2B | 75.1 | 73.8 | 52.1 | 83.4 | 70.0 | 80.9 | 60.8 |
| Qwen3-VL-Reranker-8B | 79.2 | 80.7 | 55.8 | 86.3 | 74.9 | 83.6 | 66.7 |
Implementation in Retrieval Pipelines
For optimal results, the Embedding and Reranker models are used in a two-stage retrieval process:
- Initial Recall: The Qwen3-VL-Embedding model efficiently retrieves a broad set of candidates from the vector space.
- Refined Ranking: The Qwen3-VL-Reranker model sorts these candidates using joint encoding to provide the most accurate final results.
This combined approach leverages the efficiency of the embedding model for scalability and the precision of the Reranker for fine-grained alignment.