mixedbread ai mxbai-rerank-large-v1 logo

mixedbread ai mxbai-rerank-large-v1

Visit

mxbai-rerank-large-v1 is mixedbread ai's 0.4B English reranking model under Apache 2.0, scoring 48.8 NDCG@10 on 11 BEIR datasets according to its card.

Share:
View alternatives

mixedbread ai mxbai-rerank-large-v1

mxbai-rerank-large-v1 is the largest model in mixedbread ai's first reranker family, which also includes mxbai-rerank-xsmall-v1 and mxbai-rerank-base-v1. According to its Hugging Face card, it has 0.4B parameters (F16), is tagged English, and is released under the Apache 2.0 license. The announcement blog post dates to 2024.

Status

Live on Hugging Face. The card names no successor, and the sibling models are the smaller xsmall and base versions. The card also promotes mixedbread's "Omni" multimodal, multilingual retrieval model, but does not present it as a successor.

Key Features

  • Cross-encoder reranker: scores a query against each candidate document, for use after first-stage retrieval.
  • Vendor-reported accuracy: the card reports 48.8 NDCG@10 and 74.9 Accuracy@3, aggregated over 11 BEIR datasets and evaluated with Pyserini. In the same table it lists base-v1 at 46.9 and 72.3, and bge-reranker-large at 45.2 and 70.6. These are mixedbread's own numbers.
  • Open license: Apache 2.0, so the weights can be used commercially.
  • Easy to load: works with Sentence Transformers (CrossEncoder), Transformers and Transformers.js.
  • Hosted option: mixedbread's API also offers reranking.

The card does not state a context length, so check the card before sending long documents.

Use Cases

  • Improving the order of results from BM25 or vector search in English-language search and RAG.
  • Browser or edge reranking through Transformers.js.
  • A compact open baseline against which to compare newer rerankers.

Limitation: the card tags English only, so test it before using it on other languages.

Pricing

The weights are free under Apache 2.0. Hosted API prices are not quoted here because they were not verified from mixedbread's pricing page on the date checked.

Getting Started

  1. With Sentence Transformers, load the model as a CrossEncoder and call model.rank(query, documents, top_k=...).
  2. Or load it with Transformers as AutoModelForSequenceClassification, or in the browser through Transformers.js.
  3. Feed it the top 20 to 100 candidates from your first-stage retriever and keep the highest-scoring few.

Frequently Asked Questions

What size is mxbai-rerank-large-v1?

0.4B parameters, according to the Hugging Face card.

Is it free for commercial use?

It is released under Apache 2.0.

Does it support multiple languages?

The card tags it as English. Test other languages before relying on it.

How does it compare with bge-reranker-large?

On mixedbread's own BEIR aggregate, the card reports 48.8 NDCG@10 against 45.2 for bge-reranker-large. This is a vendor-reported comparison.

Alternatives

Conclusion

mxbai-rerank-large-v1 is a small, permissively licensed English reranker that is still easy to adopt. See the model card and blog post, and compare it with newer rerankers on your own data.

Comments

No comments yet. Be the first to comment!