23 models · ranked by HuggingFace downloads
Cross-encoder reranker trained on the MS MARCO passage retrieval dataset, designed to score query-document pairs jointly rather than encoding them independently. Distilled from a 12-layer cross-encoder into 6 layers to reduce latency while retaining re-ranking accuracy. Used as a second-stage ranker on top of fast first-stage retrieval (BM25 or bi-encoder).
A 4-layer MiniLM cross-encoder fine-tuned on MS MARCO passage ranking for fast query-document relevance scoring. Offers the lowest latency in the MS MARCO cross-encoder series at the cost of some ranking accuracy.
The 12-layer variant of MiniLM fine-tuned on MS MARCO, providing the highest accuracy in the MiniLM cross-encoder family at the cost of latency. A strong default choice for reranking when compute allows.
gte-reranker-modernbert-base is a cross-encoder for text ranking. It produces fine-grained relevance judgments at the cost of higher per-pair inference latency.
jina-reranker-v2-base-multilingual is a cross-encoder for text ranking. It produces fine-grained relevance judgments at the cost of higher per-pair inference latency.
Qwen3-Reranker-0.6B ranks candidate passages against a query by fully attending to both strings in a single forward pass, making it more accurate but slower than bi-encoders.
Qwen3-Reranker-4B ranks candidate passages against a query by fully attending to both strings in a single forward pass, making it more accurate but slower than bi-encoders.
mmarco-mMiniLMv2-L12-H384-v1 is a cross-encoder for text ranking. It produces fine-grained relevance judgments at the cost of higher per-pair inference latency.
ms-marco-MiniLM-L2-v2 is a cross-encoder for text ranking. It produces fine-grained relevance judgments at the cost of higher per-pair inference latency.
Qwen3-Reranker-8B ranks candidate passages against a query by fully attending to both strings in a single forward pass, making it more accurate but slower than bi-encoders.
jina-reranker-v3 is a cross-encoder for text ranking. It produces fine-grained relevance judgments at the cost of higher per-pair inference latency.
mxbai-rerank-base-v1 ranks candidate passages against a query by fully attending to both strings in a single forward pass, making it more accurate but slower than bi-encoders.
Qwen3-VL-Reranker-8B is Qwen's multimodal cross-encoder reranker built on the Qwen3-VL-8B architecture, capable of reranking both text-only and text+image query-document pairs. It is trained with contrastive reranking objectives to score relevance between a query and a set of retrieved passages or image-text chunks, making it the first VLM-based reranker in the MTEB ecosystem.
mxbai-rerank-xsmall-v1 ranks candidate passages against a query by fully attending to both strings in a single forward pass, making it more accurate but slower than bi-encoders.
crossencoder-camembert-base-mmarcoFR is an open-source text-ranking model available on HuggingFace. Details are sourced from the public model registry.
rank1-7b is an open-source text-ranking model available on HuggingFace. Details are sourced from the public model registry.
llama-nemotron-rerank-1b-v2 ranks candidate passages against a query by fully attending to both strings in a single forward pass, making it more accurate but slower than bi-encoders.
Qwen3-VL-Reranker-2B scores query-document pairs for relevance. Used as a cross-encoder reranker, it jointly encodes the pair and outputs a single relevance score.
mxbai-rerank-large-v2 is Mixedbread AI's cross-encoder reranker built on Qwen2, covering 100+ languages. Rerankers score query-document pairs for relevance, improving retrieval quality when used after a fast bi-encoder retrieval step. Apache-2.0 licensed with text-embeddings-inference compatibility.
A sequence-classification-format wrapper of the Qwen3-Reranker-0.6B by Tom Aarsen (sentence-transformers maintainer), making it compatible with the CrossEncoder class without generative decoding. This format is faster for reranking than the generative format.
A 2-layer TinyBERT cross-encoder fine-tuned on MS MARCO passage ranking by the sentence-transformers team. Extremely small and fast, it trades accuracy for speed in reranking pipelines where a larger cross-encoder would be the throughput bottleneck.
mxbai-rerank-large-v1 is an open-source text-ranking model available on HuggingFace. Details are sourced from the public model registry.
japanese-reranker-cross-encoder-small-v1 is an open-source text-ranking model available on HuggingFace. Details are sourced from the public model registry.