BGE Reranker v2 m3

BAAI's Apache 2.0 multilingual reranker (568M parameters), by far the most downloaded one we list, built on BGE-M3.

In short. bge-reranker-v2-m3 re-sorts search results. It reads a question and a passage together and gives the pair a relevance score, so a search system can put the best passages first. It is a 568M multilingual model from BAAI under Apache 2.0, and it is by far the most downloaded reranker we list.

Summary

bge-reranker-v2-m3 is a cross-encoder from BAAI's FlagEmbedding project. Instead of turning text into vectors, it reads a query and a passage together and returns one relevance score, which the card says a sigmoid function can map to a value between 0 and 1. According to the project's news, BAAI released it on 18 March 2024 together with two LLM-based siblings, bge-reranker-v2-gemma, built on Gemma 2B, and bge-reranker-v2-minicpm-layerwise, built on MiniCPM-2B, and v2-m3 is the lightweight multilingual one. The configuration shows an XLM-RoBERTa-style model with 24 layers and 568M parameters that starts from BGE-M3, with 8,194 positions, although the card's examples cut inputs at 512 tokens.

The card gives no scores in text, only charts that rerank the top 100 results of other BGE and E5 models on BEIR, C-MTEB retrieval, MIRACL and a LlamaIndex test. Qwen's own table from June 2025, which reranks the top 100 candidates from Qwen3-Embedding-0.6B, puts it at 57.03 on MTEB-R, 72.16 on CMTEB-R, 58.36 on MMTEB-R, 59.51 on MLDR and 41.38 on MTEB-Code, against 65.80, 71.31, 66.36, 67.28 and 73.42 for Qwen3-Reranker-0.6B, so the newer 0.6B model is ahead on every column except Chinese retrieval. Those are Qwen's runs, not BAAI's.

The license is Apache 2.0, and the FP32 weights are 2.27 GB. You can run it with the FlagEmbedding package, where use_fp16 speeds up scoring at a small cost in quality, or with Transformers as a sequence-classification model. In the 30 days to 11 November 2026, Hugging Face counted about 14.7 million downloads, far more than any other reranker we list, and the repository has about 1,300 likes.

Good for

  • Multilingual reranking of the top candidates from BGE-M3 or any other first-stage retriever
  • A small, cheap cross-encoder that is easy to deploy and to fine-tune with FlagEmbedding
  • Commercial products, thanks to the Apache 2.0 license
  • A reranker with a very large user base, so examples and integrations are easy to find

Watch out for

  • The card has no scores in text, and Qwen's June 2025 table shows a newer 0.6B reranker ahead everywhere except Chinese retrieval
  • Scores are raw logits unless you normalize them, so do not compare them across models
  • The card's examples cut inputs at 512 tokens, and the 8,194 positions come only from the configuration
  • It scores one query and passage pair at a time, so cost grows with the number of candidates

Performance claims come from the developer and have not been independently verified.

How much memory does it need?

Memory figures for each variant of BGE Reranker v2 m3. Image, video, speech and music pipelines also load a text encoder and a decoder, so the memory needed can be larger than the main checkpoint.

Variants Parameters Download size Memory needed Fits on
v2-m3 (568M) FP32 568M 2.27 GB 2.61 GB Laptop or phone

Download size covers every component needed to run the model, such as the generator, text encoder and decoder. Memory needed is a stated or published measured minimum where one exists; otherwise it is the download size plus about 15%. Larger images, longer clips and higher resolutions need more.

The license in plain words

One of the most widely used permissive licenses. You may use, change and sell products built on the model with only light conditions.

More about Apache 2.0

See also

Frequently asked questions

What is BGE Reranker v2 m3?
bge-reranker-v2-m3 re-sorts search results. It reads a question and a passage together and gives the pair a relevance score, so a search system can put the best passages first. It is a 568M multilingual model from BAAI under Apache 2.0, and it is by far the most downloaded reranker we list.
Can I use BGE Reranker v2 m3 commercially?
Yes. BGE Reranker v2 m3 is released under Apache 2.0, a permissive license. You can use it commercially if you follow the standard conditions, such as keeping the license notice.
Is BGE Reranker v2 m3 open source?
BGE Reranker v2 m3 is released under Apache 2.0, a license approved by the Open Source Initiative. Training data and training code may not be public, so it is best described as open weights under an open-source license.
Can I run BGE Reranker v2 m3 on my own computer?
Yes. The smallest variant, v2-m3 (568M), needs about 2.61 GB of memory at 4-bit precision, which fits a modern laptop or a mid-range graphics card.
When was BGE Reranker v2 m3 released?
BGE Reranker v2 m3 was released by BAAI on March 18, 2024.
How long a context window does BGE Reranker v2 m3 have?
BGE Reranker v2 m3 supports a context window of up to 8K tokens.

Official links

Sources

Specifications come from these primary sources. Benchmark claims, where mentioned, are reported by the developer.

Release date and facts last verified November 11, 2026.

Related models

Model Released Parameters Context License Smallest variant needs
Qwen3-Reranker Alibaba Qwen 596M – 8.19B 32K Apache 2.0 (Permissive) 1.37 GB Laptop or phone
Jina Reranker v3.5 Jina AI 597M 128K CC BY-NC (Non-commercial) 0.46 GB Laptop or phone
BGE-M3 BAAI 568M 8K MIT (Permissive) 2.61 GB Laptop or phone
Qwen3-Embedding Alibaba Qwen 600M – 7.57B 32K Apache 2.0 (Permissive) 1.38 GB Laptop or phone