BGE Reranker v2 M3

Lightweight multilingual reranker commonly used to improve RAG result ordering after vector search.

BAAI · BGE

Editorial review

Reviewed byOpenSourcesAI EditorialLast updatedAugust 2026SourcesExact model card

Model checkpoints, context windows, provider support, local runtime compatibility, and license terms can change quickly. Verify the exact model card before production or commercial use.

Best for

RAG builders who need a practical reranker after Qdrant, Chroma, pgvector, or other vector search.

Who should use it

RAG builders who need a practical reranker after Qdrant, Chroma, pgvector, or other vector search.
Builders who want local or self-hosted testing options.
RAG builders comparing retrieval quality, latency, and multilingual coverage.

Common workflows

RAG reranking, multilingual retrieval
rag workflows
reranking workflows
multilingual workflows

Deployment and hardware notes

Much easier to run locally than frontier LLMs; CPU may be acceptable for small workloads.

License and usage notes

Apache 2.0. Open weights. Verify the exact model card and license terms for the checkpoint or hosted provider you use.

Strengths

Open weights model option for BGE workflows.
RAG builders who need a practical reranker after Qdrant, Chroma, pgvector, or other vector search.
Commonly used in local RAG stacks as a reranking step after vector search.
Tracked as RAG in the OpenSourcesAI model directory.

Limitations

Adds latency after retrieval; benchmark quality and speed on your corpus.
Much easier to run locally than frontier LLMs; CPU may be acceptable for small workloads.
Context window and limits: 8,192 tokens.
Verify the exact model card, provider docs, license, and serving support before production use.

Local workflow notes

Commonly used in local RAG stacks as a reranking step after vector search.

Local runtimes: Sentence Transformers, Transformers

Platforms: Windows, macOS, Linux

Reranker spec

Memory~2.27 GB in fp32 (568M parameters) · ~1.14 GB in fp16OutputRelevance score per query-passage pair (a cross-encoder emits no vectors)Max input8,192 tokens

Built on the bge-m3 backbone: multilingual pair scoring with an 8,192-token budget shared between query and passage, where the older BGE rerankers stop at 512. Scores pairs only — it cannot replace your embedder.