BGE-reranker-v2-m3

Overview

BGE-reranker-v2-m3 is an open-source multilingual reranking model from the Beijing Academy of Artificial Intelligence (BAAI) using transformer-based cross-encoder architecture designed specifically for reranking tasks.

Features

• Under 600 million parameters for efficient deployment
• Cross-encoder architecture for accurate relevance scoring
• Multilingual support across many languages
• Open-source under Apache 2.0 license
• Can run on consumer GPUs
• Direct similarity output without requiring embeddings

Technical Details

• Full-attention over input pairs for maximum accuracy
• Takes question and document as input
• Outputs similarity scores directly
• More accurate than bi-encoders (embedding models)
• More time-consuming than embedding models (trade-off for accuracy)

Use Cases

• Re-ranking top-k documents from embedding models
• Improving RAG system accuracy
• Two-stage retrieval pipelines
• Question answering systems
• Document relevance scoring

Performance

Strong Mean Reciprocal Rank (MRR) scores, often achieving results close to top commercial models while being open-source and self-hostable.