mxbai-rerank

mxbai-rerank is a set of open-source reranking models from Mixedbread, released under the Apache 2.0 license, with a Python package for scoring and ordering documents against a query. The v2 models (0.5B base and 1.5B large) were trained with reinforcement learning and are also available through Mixedbread's managed API.

What It Is

mxbai-rerank is a reranker library and model family used as a second stage in search and retrieval-augmented generation pipelines. You pass a query and a list of candidate documents, and the model returns them ranked by relevance. Mixedbread is described as an applied research lab focused on retrieval.

How It Works

Install the package with pip, load a model such as mxbai-rerank-base-v2 or mxbai-rerank-large-v2, and call the rank method with a query and documents. Optional settings include an instruction to steer ranking (for example, picking the best code snippet), a longer max context length (default 8k tokens, with 32k noted as compatible), and Flash Attention 2 when it is installed. The README says the models handle 100+ languages, code and SQL, and can rerank function calls for multi-tool agents.

Models and Training

The v2 line includes base (0.5B) and large (1.5B) variants, and legacy v1 models (xsmall, base, large) remain listed. The README reports benchmark results for BEIR, multilingual, Chinese and code search, and describes a three-step training process: GRPO, contrastive learning and preference learning. A technical blog post and a preprint paper describe the method.

Managed API Option

For production use, the same models can be called via the Mixedbread Python client with additional features such as object reranking and instructions. Mixedbread's broader platform adds stores, multimodal search, an MCP server and a CLI.