BGE-VL
State-of-the-art multimodal embedding model from BAAI supporting text-to-image, image-to-text, and compositional visual search. Trained on the MegaPairs dataset with over 26 million retrieval triplets.
About this tool
Overview
BGE-VL is a State-Of-The-Art multimodal embedding model that supports any visual search applications, including text-to-image, image-to-text, image&prompt-to-image, text-to-image&text, and more. Released by BAAI in March 2025.
Model Variants
Based on the MegaPairs dataset, the BAAI BGE team trained three multi-modal retrieval models:
- BGE-VL-Base
- BGE-VL-Large
- BGE-VL-MLLM
Performance
- Achieved optimal performance in 36 multi-modal embedding evaluation tasks of the Massive Multimodal Embedding Benchmark (MMEB)
- Excels in both zero-shot and supervised fine-tuning scenarios
- BGE-VL-MLLM-S1 shows 8.1% improvement on CIRCO benchmark (mAP@5) over previous state-of-the-art
- Sets new benchmark for compositional image retrieval tasks
MegaPairs Dataset
BAAI released MegaPairs, a massive synthetic dataset containing over 26 million multimodal retrieval instruction-tuning triplets that powers BGE-VL.
License
Released under MIT license - completely free for both academic and commercial use.
Availability
Models and documentation available on GitHub, Hugging Face, and the official BGE documentation site.
Loading more......
Information
Categories
Tags
Similar Products
6 result(s)