mxbai-rerank-v3.1-listwise

mxbai-rerank-v3.1-listwise is now available as the default reranker.

It delivers gpt-5.6-sol-level ranking quality with substantially improved latency over mxbai-rerank-v3-listwise. It is excellent at complex tasks like recency-aware ranking, source-priority resolution, and multi-step composite instructions, compared with leading LLMs and pointwise rerankers.

Quality versus latency on ViDoRe v3: mxbai-rerank-v3.1-listwise reaches gpt-5.6-sol quality at roughly 61 times lower latency, while Cohere and the GPT-5.6 rerankers are slower.

Sampled queries per ViDoRe subset with Mixedbread base search (Wholembed v3) as first stage.

Show the data behind this chart

Reranker on top of the first stage	NDCG@10 (ViDoRe v3, sampled queries)	Latency per query
First stage only (Wholembed v3)	0.62	0.10 s
+ cohere rerank-4-pro	0.66	1.6 s
+ gpt-5.6-terra (high)	0.70	36 s
+ gpt-5.6-sol (high)	0.70	40 s
+ gpt-5.6-luna (high)	0.70	73 s
+ mxbai-rerank-v3.1-listwise	0.71	0.65 s

Higher NDCG@10 and lower latency are better. Values are decoded from the chart's plotted positions (NDCG@10 to ±0.005, latency to about ±5%); the 61× figure is gpt-5.6-sol's latency divided by v3.1's.

Unlike pointwise rerankers measuring the relevance of individual documents, our listwise v3 reads the whole candidate set for better reasoning at the cost of higher ranking latency.

Now, by rewriting its underlying inference engine, v3.1 reranks faster across every input size, matching pointwise rerankers. In production, the speedup ranges from ~25% on typical queries to ~54% on long-tail inputs (64–128k tokens).

Median reranking latency by input size: mxbai-rerank-v3.1-listwise is faster than v3 at every input size.

Show the data behind this chart

Input size (tokens)	v3 median latency	v3.1 median latency	Speedup
< 8k	159 ms	119 ms	25%
8–16k	258 ms	188 ms	27%
16–32k	380 ms	243 ms	36%
32–64k	611 ms	348 ms	43%

Median reranking latency in production; lower is better. Values are decoded from the chart's plotted positions (±3 ms). The ~54% speedup quoted in the text is for 64–128k-token inputs, which the chart does not show.

For better threshold cutoffs, we also changed v3.1 to return content-dependent relevance scores instead of a fixed rank-based ladder in v3.

mxbai-rerank-v3.1-listwise is available today through Mixedbread Search:

from mixedbread import Mixedbread

client = Mixedbread()

results = client.stores.search(
    store_identifiers=["my-store"],
    query="when is my flight to London? The most recent valid booking wins",
    search_options={
      "rerank": {
          "model": "mixedbread-ai/mxbai-rerank-v3.1-listwise",
      }
    },
)