Curiosity · AI Model
BAAI BGE Reranker v2-M3
BGE Reranker v2-M3 is BAAI's open-weight multilingual reranker, purpose-built to pair with BGE-M3 embeddings. It is one of the strongest self-hostable rerankers — runs on a single consumer GPU and powers the reranking stage in many OSS RAG templates (LlamaIndex, Haystack, LangChain).
Model specs
- Vendor
- BAAI
- Family
- BGE Reranker
- Released
- 2024-03
- Context window
- 8,192 tokens
- Modalities
- text
Strengths
- MIT-style open licence — fully self-hostable
- Natural pair with BGE-M3 embeddings
- Strong MIRACL multilingual scores
- Runs comfortably on a single consumer-grade GPU
Limitations
- Self-hosting requires ops investment vs a managed Rerank API
- Slightly trails Cohere Rerank 3 on English-only retrieval
- Large reranker variants (bge-reranker-v2.5-gemma2) need more VRAM
Use cases
- Second-stage reranking in open-source RAG pipelines
- Hybrid retrieval paired with BGE-M3 dense+sparse output
- Multilingual enterprise search in regulated environments
- On-prem research and academic retrieval systems
Benchmarks
| Benchmark | Score | As of |
|---|---|---|
| MIRACL NDCG@10 | ≈72 | 2024 |
| BEIR NDCG@10 uplift | +10 | 2024 |
Frequently asked questions
What is BGE Reranker v2-M3?
BGE Reranker v2-M3 is an open-weight multilingual cross-encoder reranker from BAAI. It reorders candidate passages produced by a first-stage retriever such as BGE-M3 embeddings or BM25 for higher top-k precision in RAG.
How do I use it with BGE-M3?
A typical pipeline retrieves top-100 candidates with BGE-M3 dense embeddings (optionally combined with BGE-M3 sparse scores), then reranks those 100 candidates with BGE Reranker v2-M3, keeping the top 5 for the LLM context.
Is it free to use?
Yes — BGE Reranker v2-M3 is published on Hugging Face under an open licence and can be used commercially. You self-host on your own GPU or CPU.
How much VRAM does it need?
The base model runs comfortably on a 12–16 GB consumer GPU. Larger variants (bge-reranker-v2.5-gemma2-lightweight or similar) require more VRAM but deliver higher MIRACL scores.
Sources
- Hugging Face — BAAI/bge-reranker-v2-m3 — accessed 2026-04-20
- BAAI FlagEmbedding repo — accessed 2026-04-20