GME Qwen2-VL 2B

GME-Qwen2-VL-2B-Instruct is a multimodal embedding model that produces unified vector representations from text, image, or combined image-text inputs, enabling any-to-any retrieval across modalities. Developed by Tongyi Lab at Alibaba Group, it is based on the Qwen2-VL multimodal large language model and supports dynamic image resolution with a maximum sequence length of 32,768 tokens and an embedding dimension of 1,536.

Key Capabilities

The model processes three input types—text, image, and image-text pairs—and outputs a single embedding that can be used for text retrieval, image-to-text search, image-to-image search, and fused-modal search. This makes it suitable for tasks such as multimodal retrieval-augmented generation (RAG) on documents, visual document retrieval, and cross-modal search.

Benchmark Performance

On the Universal Multimodal Retrieval Benchmark (UMRB), the 2.2B-parameter model achieves an average score of 64.45, outperforming larger models such as DSE (4.2B, 50.04) and E5-V (8.4B, 42.52). It scores 87.84 on text-to-visual-document retrieval and 76.47 on text-to-image-text retrieval. On the MTEB English leaderboard, it scores 65.27, and on MTEB Chinese, 66.92. The paper describing the model has been accepted to CVPR 2025.

Strengths

• Unified representation: A single model handles text-only, image-only, and fused inputs without separate encoders.
• Strong visual document understanding: Excels at retrieving document screenshots and academic papers, critical for multimodal RAG.
• Dynamic resolution: Adapts to varying image sizes without fixed resizing.

Limitations

The model accepts only single-image inputs and is trained on English data only; multilingual multimodal performance is not guaranteed.