BGE Reranker
Beijing Academy of Artificial Intelligence (BAAI)BAAI's cross-encoder reranking line within the BGE/FlagEmbedding project, separate from the BGE bi-encoder embedders. bge-reranker-v2-m3 is the workhorse - a multilingual 0.6B model built on bge-m3 - and the line also ships the earlier XLM-R base and large rerankers plus LLM-based v2-gemma and layerwise v2-minicpm variants that trade latency for accuracy. It is the second most downloaded reranker family on the Hub.
Added 2026-09-11 as a seventh product in this research set, not on the original list. It surfaced because BAAI tags its rerankers `text-classification` rather than `text-ranking`, so the category's pipeline-tag sweep missed a family drawing roughly 25M downloads a month - larger than every reranker here except the ms-marco cross-encoders. Distinct from the BGE embedding line, which is a different product. Verified 2026-09-11 via the bge-reranker-v2-m3 and bge-reranker-v2.5-gemma2-lightweight model cards and the FlagEmbedding repository.
Openness
3 high confidence- weights
- open(ungated safetensors for every checkpoint in the line)
- data
- closed(no reranker training data released
- code
- open(FlagEmbedding ships the reranker fine-tuning pipeline, configs and data format)
- license
- Apache-2.0(OSI
The line ships under three licences and multi_sku_rule resolves it on the most restrictive DISTRIBUTED SKU: bge-reranker-v2.5-gemma2-lightweight carries the Gemma licence, which the ladder places at use_bounded, so the family lands at 3 rather than at the Apache and MIT rung its two most-downloaded checkpoints would reach alone. Note the outcome is the same either way here, since no reranker training corpus is released.
- https://huggingface.co/BAAI/bge-reranker-v2-m3/raw/main/README.md recorded 2026-09-11
Front matter for bge-reranker-v2-m3 reads license: apache-2.0; the card lists the family - base and large on XLM-R, v2-m3 on bge-m3, v2-gemma on gemma-2b, v2-minicpm-layerwise - and points at FlagOpen/FlagEmbedding, with a Fine-tune section giving the training data format and run commands but naming no released corpus.
- https://huggingface.co/BAAI/bge-reranker-v2.5-gemma2-lightweight/raw/main/README.md recorded 2026-09-11
Front matter for bge-reranker-v2.5-gemma2-lightweight reads license: gemma, which is the most restrictive licence among the distributed checkpoints in this line.
- https://raw.githubusercontent.com/FlagOpen/FlagEmbedding/master/README.md recorded 2026-09-11
FlagEmbedding README links examples/finetune/reranker and examples/finetune/embedder as the training pipelines and a dataset directory, and records the 2023 release of the BGE massive training data for the embedding models.
Adoption
5 high confidenceAbout 24.9M downloads in the trailing 30 days across the line - v2-m3 18.2M, base 3.6M, large 2.6M, v2-gemma 311K, v2.5-gemma2-lightweight 164K, v2-minicpm-layerwise 6K. v2-m3 alone clears the >10M band. Community mirrors and GGUF/ONNX conversions are excluded.
- https://huggingface.co/api/models/BAAI/bge-reranker-v2-m3 recorded 2026-09-11
downloads: 18212308; gated: false; private: false; license apache-2.0 for BAAI/bge-reranker-v2-m3.
Capability
3 medium confidenceRung 3, solid workhorse: the most widely deployed multilingual reranker of the 2024 generation, with the strongest MIRACL number on the comparison table but a BEIR average and a code-retrieval score well behind the 2025-26 frontier. One rung below jina-reranker, which is measured 61.94 to 56.51 on the same table. Confidence is medium because the figures come from a competitor published comparison rather than from BAAI own card, which publishes no headline BEIR number.
- https://huggingface.co/jinaai/jina-reranker-v3/raw/main/README.md recorded 2026-09-11
Comparison table lists bge-reranker-v2-m3 at 0.6B with 56.51 BEIR, 69.32 MIRACL - the highest MIRACL value in the table - 67.88 MKQA and 36.28 CoIR, against jina-reranker-v3 at 61.94 BEIR.
- https://huggingface.co/BAAI/bge-reranker-v2-m3/raw/main/README.md recorded 2026-09-11
Model List describes bge-reranker-v2-m3 as a lightweight multilingual reranker built on bge-m3 and recommends it for multilingual, Chinese and English, and efficiency scenarios.
Verified 2026-09-11