May 2026 Summaries

Mixedbread has released mxbai-rerank-v3-listwise in preview through Mixedbread Search, positioning it as a listwise reranking model designed alongside Wholembed v3 to improve strong initial retrieval results where pointwise rerankers may provide little benefit or reduce quality. Unlike models that score documents independently, it evaluates candidate documents collectively to resolve relationships such as amendments, superseding updates, source authority, and recency, while accepting natural-language ranking instructions. On the 56-run multilingual, multidomain ViDoRe v3 benchmark, Wholembed v3’s average NDCG@10 increased from 0.603 to 0.669, an average gain of 10.92%, with larger improvements on difficult German industrial and French HR subsets. In a separate 900-example instruction-following evaluation, the model achieved 0.93 MRR and 88.6% Accuracy@1, outperforming several pointwise competitors, particularly on tasks requiring recency and source-priority reasoning.