Qwen3-VL-Embedding and Qwen3-VL-Reranker: Models for Next-Generation Multimodal Search

Qwen has released Qwen3-VL-Embedding and Qwen3-VL-Reranker. Both models are built on top of Qwen3-VL and are designed to handle text, images, screenshots, and video within a single framework.

The core is a two-stage retrieval structure. The Embedding model converts inputs into vectors in a shared semantic space to quickly find candidates, while the Reranker encodes (Query, Document) pairs together to produce more precise relevance scores.

• Embedding: a dual-tower structure that independently encodes single- or mixed-modal inputs
• Reranker: a single-tower structure that processes Query and Document together, using cross-attention
• Output method: Embedding produces vectors, while Reranker computes scores from the generation probability of yes/no tokens

The models come in two specs, 2B and 8B. The Embedding models provide vectors of 2048 and 4096 dimensions respectively, and support a 32K sequence length, quantization support, MRL support, and instruction-aware capability.

In evaluations, Qwen3-VL-Embedding-8B outperformed previous open-source models and even commercial services on MMEB-v2, showing particular strength in image, visual document, and video retrieval subtasks. However, on the text-only MMTEB, it performed lower than the text-only Qwen3-Embedding of the same size, though the authors note it remains competitive compared to models of similar scale.

On the Reranker side, all models consistently outperformed the base embedding models and existing rerankers, with the 8B model achieving the best performance on most tasks. The authors emphasize that this combination is a practical solution for strengthening both initial recall and precise reranking in large-scale multimodal search.