MS MARCO Cross-Encoders
UKP Lab (TU Darmstadt)The `cross-encoder/ms-marco-*` family of BERT and MiniLM cross-encoders trained on MS MARCO Passage Ranking and published alongside SentenceTransformers. Each checkpoint takes a (query, passage) pair and returns a single relevance logit, and the family spans TinyBERT-L2 through MiniLM-L12 so a deployment can trade throughput against NDCG. It is the default self-hosted reranker in most retrieve-and-rerank stacks.
Added 2026-09-11 for the embeddings_retrieval promotion. The family is old - the v2 checkpoints date to 2022 - and English-only, but it remains by a wide margin the most downloaded reranker line on the Hub.
Openness
5 high confidence- weights
- open(Apache-2.0 ungated safetensors for every checkpoint in the family)
- data
- open(MS MARCO Passage Ranking
- code
- open(SentenceTransformers ships the MS MARCO cross-encoder training scripts the card links to)
- license
- Apache-2.0(OSI)
The whole chain is public: an Apache-2.0 ungated checkpoint, a training corpus named in the card's front matter, and the actual training scripts that produced it, shipped in the SentenceTransformers examples tree. That is a fully open release earned on the evidence rather than assumed from reputation.
- https://huggingface.co/cross-encoder/ms-marco-MiniLM-L6-v2/raw/main/README.md recorded 2026-10-20
Front matter reads license: apache-2.0 and datasets: sentence-transformers/msmarco; the body says the model was trained on MS MARCO Passage Ranking and links the SBERT.net MS MARCO training code.
- https://raw.githubusercontent.com/UKPLab/sentence-transformers/master/examples/cross_encoder/training/ms_marco/README.md recorded 2026-10-20
The cross_encoder/training/ms_marco directory of SentenceTransformers, listing training_ms_marco_bce.py, training_ms_marco_cmnrl.py, training_ms_marco_listnet.py and training_ms_marco_lambda.py as the scripts that train a CrossEncoder on MS MARCO.
Adoption
5 high confidenceAdoption is measured as Hugging Face downloads summed across the cross-encoder/ms-marco-* checkpoints, most of it for MiniLM-L6-v2 with a smaller share for MiniLM-L4-v2.
- https://huggingface.co/api/models/cross-encoder/ms-marco-MiniLM-L4-v2 recorded 2026-09-11
downloads: 5701227; gated: false; private: false for cross-encoder/ms-marco-MiniLM-L4-v2.
- https://huggingface.co/api/models/cross-encoder/ms-marco-MiniLM-L6-v2 recorded 2026-09-11
downloads: 87478105; gated: false; private: false; license apache-2.0 for cross-encoder/ms-marco-MiniLM-L6-v2.
Capability
2 high confidenceThe v2 checkpoints are an older generation, cover English only, and cap at 512 tokens, which is a narrow and dated scope rather than a comment on quality. Its published numbers are credible and it remains the default self-hosted reranker, but popularity isn't what the score measures; it sits one step below bge-reranker, which is multilingual, longer-context and a generation newer.
- https://huggingface.co/cross-encoder/ms-marco-MiniLM-L6-v2/raw/main/README.md recorded 2026-09-11
Performance table gives cross-encoder/ms-marco-MiniLM-L6-v2 at 74.30 NDCG@10 on TREC DL 19 and 39.01 MRR@10 on MS MARCO dev, with the family spanning TinyBERT-L2 at 69.84 to MiniLM-L12 at 74.31; language front matter is en.
Verified 2026-09-11