Qwen releases multimodal embedding and reranker models
Alibaba's Qwen team has released Qwen3-VL-Embedding and Qwen3-VL-Reranker, extending its retrieval models from text to images, screenshots and video, and says the 8B embedding model tops the MMEB-V2 benchmark.
The Qwen team announced two new model series for multimodal retrieval: Qwen3-VL-Embedding and Qwen3-VL-Reranker. Both are built on its Qwen3-VL foundation models. They follow the text-only Qwen3-Embedding and Qwen3-ReRanker models, which the team open-sourced in June 2025.
Each series comes in 2B (28 layers) and 8B (36 layers) sizes, and all four models have a 32K sequence length. The embedding models output 2048-dimensional vectors at 2B and 4096-dimensional vectors at 8B. They support quantized embeddings and Matryoshka-style (MRL) user-specified dimensions. All four models accept task-specific instructions, and the team says they support more than 30 languages.
The embedding model uses a dual-tower design. It accepts text, images, screenshots, video or mixed inputs and takes the last-layer hidden state at the [EOS] token as the representation. The reranker is single-tower. It encodes a query-document pair together and produces a relevance score from the predicted probabilities of "yes" and "no" tokens. Qwen presents the two as a pipeline: the embedding model handles initial recall and the reranker refines the results.
On MMEB-V2, Qwen says Qwen3-VL-Embedding-8B achieves state-of-the-art results, ahead of all previous open-source models and proprietary services. The team reports top results on the image, visual document and video retrieval subtasks. On the text-only multilingual MMTEB benchmark, the model trails the text-only Qwen3-Embedding model of similar size. The excerpt of the post cuts off before its comparison with other models.
The excerpt does not include specific benchmark scores, licensing terms or download locations.