Alibaba's Qwen3-VL-Embedding Hits #1 Ranking by Unifying Text, Images, and Video Search
Alibaba's Qwen3-VL-Embedding brings text, image, video, and document search into a single open-source embedding model, topping the MMEB-V2 leaderboard at 77.8
Takeaways
- Alibaba releases Qwen3-VL-Embedding and Qwen3-VL-Reranker, open-source models for multimodal retrieval across text, images, video, and documents.
- Qwen3-VL-Embedding-8B scores 77.8 on MMEB-V2, ranking #1 among all open-source and proprietary multimodal embedding models.
- Four weights released (2B and 8B for both embedding and reranking) under Apache 2.0, available on Hugging Face and ModelScope.
- Supports Matryoshka embeddings (flexible dimensions), quantized embeddings, 30+ languages, and instruction-aware task customization.
- Two-stage pipeline: embedding model for broad recall, reranker for precision re-scoring; 8B reranker scores 79.2 on MMEB-V2 retrieval.
- vLLM >= 0.14.0 supported for production-scale batched inference; includes end-to-end multimodal RAG example notebook.
Qwen3-VL-Embedding is Alibaba's new open-source suite for multimodal retrieval, and it does something the embedding world has been waiting for: a single model that can embed text, images, screenshots, videos, and any mix of those together into one shared vector space. Pair it with the companion Qwen3-VL-Reranker and you have a complete, end-to-end retrieval pipeline that handles any modality combination from query to final ranking.
The problem it actually solves
Most production RAG systems today are text-only. When you want to retrieve from a PDF screenshot, a product image, or a video clip, you either need separate specialized models per modality or you lose the visual signal entirely. There has been no clean way to run a multimodal embedding model that takes both images and text as input and produces a single embedding vector -- a fundamentally different use case from VLM generation, covering retrieval, RAG with images and documents, cross-modal search, and clustering. Qwen3-VL-Embedding closes that gap.
What's in the box
The release introduces the Qwen3-VL-Embedding and Qwen3-VL-Reranker model series, built on the Qwen3-VL foundation model, providing an end-to-end pipeline for high-precision multimodal search by mapping diverse modalities -- text, images, document images, and video -- into a unified representation space. Four model weights are released in total:
- Qwen3-VL-Embedding-2B -- 28 layers, 2048-dim embeddings, 32K context
- Qwen3-VL-Embedding-8B -- 36 layers, 4096-dim embeddings, 32K context
- Qwen3-VL-Reranker-2B -- 28 layers, relevance scoring
- Qwen3-VL-Reranker-8B -- 36 layers, best overall reranking performance
Both model series inherit the multilingual capabilities of Qwen3-VL, supporting more than 30 languages, and are released in 2B and 8B parameter sizes to accommodate diverse deployment requirements. All four weights are available on Hugging Face and ModelScope under the Apache 2.0 license, meaning commercial use is permitted.
Two architectures, two jobs
The embedding and reranking models are architecturally distinct by design, and understanding the split matters for how you'd use them.