What it is
mxbai-rerank v2 is a pair of open rerankers from Mixedbread, released in March 2025: mxbai-rerank-base-v2 (0.5 billion parameters) and mxbai-rerank-large-v2 (1.5 billion). A reranker rescores the shortlist a search system has already retrieved by reading the query and each document together, which is slower than comparing vectors but more precise. Both score one document at a time (pointwise), are built on Qwen2.5 and were trained with reinforcement, contrastive and preference learning.
Mixedbread says they cover more than 100 languages and code, with 8,000 tokens of context by default and compatibility up to 32,000, against 512 for v1. The weights are on Hugging Face under Apache 2.0, with a small mxbai-rerank Python package.
Mixedbread's newer reranker, mxbai-rerank-v3.1-listwise, is proprietary and available only inside its hosted Mixedbread Search, where it is now the default and large-v2 remains a pointwise option. The docs describe no standalone rerank endpoint; reranking adds $3.50 per 1,000 semantic searches.
Our review
mxbai-rerank v2 earns its place on licence and size rather than current accuracy. Both models are Apache 2.0 with no commercial strings, which matters now that Jina's newer reranker needs a paid Elastic licence for commercial self-hosting, and the 0.5B base model is small enough to run beside an application on a modest GPU. Mixedbread's own timing on one A100 puts large-v2 at 0.89 seconds and base-v2 at 0.67 seconds per query on NFCorpus, against 3.05 seconds for bge-reranker-v2-m3 and 7.20 for bge-reranker-v2-gemma.
At launch it led the open field on English. Mixedbread reports a BEIR average of 57.49 NDCG@10 for large-v2 and 55.57 for base-v2, against 55.39 for Cohere rerank-3.5 and 54.54 for Voyage rerank-2, plus 39.2 on ToolRet tool retrieval, a narrow win over the previous best of 39.09. Both BGE rerankers beat it on multilingual Mr.TyDi (30.99 and 30.40 against 29.79), so the 100-language claim deserves testing on your own data. All of these are Mixedbread's measurements.
Time has moved on. Every commercial rival in that table has been replaced, and Jina reports 63.20 on BEIR for its 0.6B Jina Reranker v3.5, though with a different setup, so the numbers are not directly comparable. Mixedbread's effort has gone into the proprietary listwise v3.1, which it reports matching a frontier LLM reranker on ViDoRe v3 at 0.65 seconds per query, but which runs only inside Mixedbread Search. The open GitHub repository was last updated in September 2025, and the Python package is still at version 0.1.6.
Hosted, the economics change. Mixedbread charges $7.50 per 1,000 semantic searches with reranking, $3.50 of it for the rerank step, so a million reranked queries a month is $7,500 before indexing and storage. At Mixedbread's own 0.89 seconds a query on an A100, the same million queries is about 250 GPU-hours of large-v2, well within one card's month.
Where it earns its place
A search engineer at a 20-person SaaS company running BM25 in Elasticsearch
Improve the relevance of in-app help search without adding a vendor
Put base-v2 behind the existing BM25 shortlist on a small GPU; in Mixedbread's tests BM25 plus base-v2 averaged 55.57 on BEIR. Apache 2.0 avoids a licence review, and the only running cost is the GPU.
A developer building an agent with hundreds of tools
Choose which tool descriptions to show the model for each request
Rerank tool descriptions against the request with large-v2, which scored 39.2 on ToolRet with BM25 as the first stage against 31.97 for BM25 alone. Pointwise scoring keeps each tool's score independent of the others.
A head of knowledge management at a law firm
Decide between a hosted search service and self-hosting a reranker
Mixedbread Search with reranking costs $7.50 per 1,000 queries and defaults to the closed v3.1 model; self-hosting v2 keeps documents on firm hardware but means owning the GPU, monitoring and updates.
Before you commit
- Setup
- pip install mxbai-rerank, load mixedbread-ai/mxbai-rerank-base-v2 or large-v2 from Hugging Face, and pass a query with its candidate documents.
- Integrations
- The mxbai-rerank Python package and Transformers; inside Mixedbread Search, large-v2 is selectable in search_options.rerank with top_k and metadata controls.
- Data and privacy
- Self-hosted under Apache 2.0, so documents never leave your infrastructure. Mixedbread Search offers bring-your-own-bucket on Enterprise; SOC 2 status is not published on its pricing page.
- Cost at scale
- Self-hosted, a million queries at Mixedbread's 0.89 seconds each is about 250 A100-hours. Hosted, Mixedbread Search charges $7.50 per 1,000 reranked semantic searches, or $7,500 per million, plus $1.50 to $3 per 1M tokens indexed.
- Limits
- 8K tokens per query-document pair by default, 32K compatible; pointwise, so cost grows linearly with candidates; Mixedbread Search Starter allows 100 queries a minute and Scale 2,400.
What's good
- Apache 2.0, no commercial restrictions
- 0.67 to 0.89 seconds a query on one A100 in Mixedbread's tests
- 8K-token context, 32K compatible
- Beats BGE on Mixedbread's code and tool-retrieval tests
What's not
- Released March 2025; open repository idle since September 2025
- Trails BGE on multilingual Mr.TyDi
- Newer v3.1 listwise model is closed and tied to Mixedbread Search
- Benchmarks are Mixedbread's own
- No standalone hosted rerank API
Best for
Teams that want a small, permissively licensed reranker to run on their own hardware.
Alternatives
mxbai-rerank v2: common questions
- Is mxbai-rerank v2 worth it?
- Still the safest small reranker to self-host on licence alone, though its benchmark lead dates from March 2025. We rate it 3.8 out of 5.
- How much does mxbai-rerank v2 cost?
- Free under Apache 2.0. Hosted only inside Mixedbread Search: semantic search $4 per 1,000 queries, $7.50 with reranking; Scale plan $20/month in credits.
- What are the downsides of mxbai-rerank v2?
- Released March 2025; open repository idle since September 2025; trails BGE on multilingual Mr.TyDi; newer v3.1 listwise model is closed and tied to Mixedbread Search; benchmarks are Mixedbread's own; no standalone hosted rerank API.
- Who is mxbai-rerank v2 best for?
- Teams that want a small, permissively licensed reranker to run on their own hardware.
- How much does mxbai-rerank v2 cost for a team or in production?
- Self-hosted, a million queries at Mixedbread's 0.89 seconds each is about 250 A100-hours. Hosted, Mixedbread Search charges $7.50 per 1,000 reranked semantic searches, or $7,500 per million, plus $1.50 to $3 per 1M tokens indexed.
- Is mxbai-rerank v2 safe for company data?
- Self-hosted under Apache 2.0, so documents never leave your infrastructure. Mixedbread Search offers bring-your-own-bucket on Enterprise; SOC 2 status is not published on its pricing page.
- What are the best mxbai-rerank v2 alternatives?
- The closest alternatives we have reviewed are Qwen3-Embedding and Qwen3-Reranker, zerank-2, Jina Reranker v3.5 and Voyage rerank-3.
Answers checked Oct 22, 2026.