Mixedbread Releases mxbai-Rerank-v3.1-Listwise with 61× Performance Gains
Key Takeaways
- ▸v3.1 achieves GPT-5.6-sol quality ranking while running 61× faster than that baseline
- ▸25–54% latency improvement over v3, with especially strong gains on long-tail queries (64–128K tokens)
- ▸Listwise approach maintains superior reasoning while matching pointwise reranker speed via optimized inference
Summary
Mixedbread has released mxbai-rerank-v3.1-listwise, an upgraded version of its listwise reranker offering substantial performance improvements over its predecessor. The new version delivers GPT-5.6-sol-level ranking quality while being significantly faster, with a rewritten inference engine that improves latency by 25–54% across all input sizes, from typical 8K-token queries to long-tail 128K-token workloads.
The model excels at sophisticated ranking tasks including recency-aware ranking, source-priority resolution, and multi-step composite instructions. Unlike pointwise rerankers that evaluate individual documents in isolation, the listwise approach reads the entire candidate set for more nuanced reasoning—and v3.1 achieves this without the latency penalty of previous versions.
Additionally, Mixedbread refined v3.1's relevance scoring, shifting from v3's fixed rank-based ladder to content-dependent scores for improved threshold calibration. The updated model is now available as the default reranker through Mixedbread Search.
- Enhanced for complex retrieval tasks: recency filtering, multi-priority resolution, and composite instructions
