March 2025 Summaries

Mixedbread AI has released the Apache 2.0-licensed mxbai-rerank-v2 family of open-source reranking models, including a 0.5B-parameter base model and a 1.5B-parameter large model, designed to improve search systems by reordering initially retrieved results according to deeper query-document relevance. Built on Qwen-2.5 and trained through reinforcement learning, contrastive learning, and preference learning, the models support more than 100 languages, contexts up to 8,000 tokens with 32,000-token compatibility, and uses beyond document retrieval such as code search, tool selection, e-commerce discovery, and structured JSON data. Mixedbread reports that its large model achieved a leading 57.49 NDCG@10 score on the BEIR English retrieval benchmark, while the base model scored 55.57, both exceeding several cited open- and closed-source competitors using BM25 as the initial retriever. The company also reports favorable latency results on an A100 GPU, with the large v2 model processing queries substantially faster than some larger competing rerankers, and offers access through Hugging Face, a Python package, and its API.