Curiosity · AI Model
mxbai-rerank-large-v1
mxbai-rerank-large-v1 is an open cross-encoder reranking model from mixedbread.ai, released April 2024. It was the first open reranker to match or beat Cohere Rerank 2 on the BEIR retrieval benchmark while being released under Apache 2.0 on Hugging Face. The 'large' variant (~435M parameters) sits alongside base (184M) and xsmall (~70M) siblings, letting teams choose their quality/latency trade-off. Widely used as a local alternative to Cohere Rerank for on-prem RAG and air-gapped deployments.
Model specs
- Vendor
- mixedbread.ai
- Family
- mxbai-rerank
- Released
- 2024-04
- Context window
- 512 tokens
- Modalities
- text
Strengths
- Apache 2.0 licence
- Matches commercial rerankers on BEIR
- Three sizes for quality/latency trade-off
- Easy integration via sentence-transformers
Limitations
- Short 512-token context per document
- English-heavy training — weaker on multilingual vs Cohere Rerank 3
- Requires GPU at scale for good throughput
- No long-document support without chunking
Use cases
- Open-source RAG reranking
- On-prem or air-gapped search
- Cost-sensitive retrieval pipelines
- Hybrid BM25 + dense + rerank stacks
Benchmarks
| Benchmark | Score | As of |
|---|---|---|
| BEIR average NDCG@10 | matched/edged Cohere Rerank 2 at release | 2024-04 |
| MS MARCO dev MRR@10 | ≈0.44 (large) | 2024-04 |
Frequently asked questions
What is mxbai-rerank-large-v1?
An open cross-encoder reranker from mixedbread.ai, released April 2024 under Apache 2.0. The 'large' variant has ~435M parameters.
How does it compare to Cohere Rerank?
On English BEIR benchmarks it matches or slightly exceeds Cohere Rerank 2. Cohere Rerank 3 Multilingual is stronger on multilingual workloads; mxbai is stronger on cost and openness.
How do I deploy mxbai-rerank?
As a standard cross-encoder via sentence-transformers or Hugging Face transformers. A lightweight Python service plus a GPU suffices.
Sources
- mxbai-rerank-large-v1 on Hugging Face — accessed 2026-04-20
- mixedbread.ai blog — mxbai-rerank — accessed 2026-04-20