MXBAI Reranker Large V2
Deploy MXBAI Reranker Large V2 on a dedicated GPU from $0.20 per hour, billed per minute, on 1× RTX A4000. Mixedbread's 1.5B second-generation reranker, trained with reinforcement and preference learning for multilingual text, code and long-context retrieval while remaining substantially smaller than many LLM-based rerankers.
HexGrid Cloud price
$0.20/hr
1× RTX A4000
Context
8K
32K compatible
Parameters
1.54B
ProRank-1.5B
Languages
100+
109 HF metadata
Running MXBAI Reranker Large V2 on HexGrid Cloud
What you get when you deploy with us, beyond the hourly rate.
Single tenant by default
The model runs on a GPU that is yours for the life of the instance. No shared endpoint, no queue behind other tenants.
Per-minute billing
You pay for the GPU, not the token. Stop the instance and billing stops with it.
OpenAI-compatible endpoint
Point an existing SDK at your instance by changing the base URL. No rewrite required.
Your weights, your data
Prompts and outputs stay inside your instance. Nothing is logged, sampled or used for training.
MXBAI Reranker Large V2 GPU sizing and cost
What the model needs at each precision, and the cheapest HexGrid Cloud configuration that holds it.
| Precision | vRAM needed | Runs on | Price |
|---|---|---|---|
BF16 Recommended | 4 GB | 1× RTX A4000 | $0.20/hr |
FP8 | 2 GB | 1× RTX A4000 | $0.20/hr |
INT4 | 1 GB | 1× RTX A4000 | $0.20/hr |
Mixedbread reports 0.89 seconds average latency per query on NFCorpus using an NVIDIA A100 80GB. That benchmark is workload-specific, so it should be surfaced separately from generic throughput. Measure candidate pairs/s and queries/s on HexGrid's exact GPU, max length and batch settings before publishing deployment performance.
About MXBAI Reranker Large V2
What MXBAI Reranker Large V2 is built for, and where it falls short.
MXBAI Reranker Large V2, also called ProRank-1.5B, is the larger checkpoint in Mixedbread's second-generation open reranker family. It jointly scores query-document pairs and is intended as a second-stage ranker after BM25, dense retrieval or hybrid search.
The model uses a Qwen2-family causal-language-model architecture and Mixedbread describes the V2 family as being initialized from Qwen 2.5. Its checkpoint contains roughly 1.54B parameters, 28 transformer layers and a 32K positional limit.
Mixedbread trained the model in three stages: GRPO to establish relevance-decision behavior, contrastive learning to sharpen query-document relationships, and preference learning to improve ranking order.
For normal use Mixedbread describes V2 as supporting long contexts up to 8K tokens while remaining compatible with 32K. The official mxbai-rerank library defaults to 8192 and can be configured up to the model's 32K limit.
What people run it for
RAG reranking
Rescore retrieved chunks before sending the highest-ranked context to a language model.
Multilingual search
Improve relevance across global document collections with support for more than 100 languages.
Code search
Rerank candidate functions, snippets, files or documentation returned by lexical or semantic code retrieval.
Hybrid search
Use BM25, vector retrieval or a hybrid system for recall, then apply the reranker as the semantic precision stage.
Strengths
- Strong publisher-reported BEIR average of 57.49 using BM25 first-stage retrieval
- 100+ language support with particularly strong published Chinese retrieval results
- Code-search support in addition to conventional document retrieval
- 32K architectural context with 8K production-oriented usage documented by Mixedbread
- Three-stage GRPO, contrastive and preference training is specifically optimized for reranking rather than general text generation
- Apache 2.0 licence
- Compatible with Sentence Transformers CrossEncoder and Mixedbread's dedicated mxbai-rerank package
Limitations
- Like other cross-encoder rerankers, inference cost increases with every candidate document sent for reranking
- The 32K architectural maximum should not be interpreted as the default production pair length; Mixedbread describes the model as 8K long-context with 32K compatibility
- Publisher latency numbers are tied to a specific A100 80GB benchmark on NFCorpus and are not universal deployment guarantees
- Scores across BEIR, multilingual, Chinese and code-search columns come from different benchmark collections and should not be compared as though they share the same scale
- The model is designed as a second-stage scorer rather than an embedding model for precomputing a corpus index
Quickstart
Use Sentence Transformers CrossEncoder or Mixedbread's mxbai-rerank package. A model-specific HexGrid HTTP serving contract was not found in the public docs.
from sentence_transformers import CrossEncoder
model = CrossEncoder('mixedbread-ai/mxbai-rerank-large-v2')
query = 'How does retrieval augmented generation work?'
documents = [
'Retrieval-augmented generation retrieves external information before producing an answer.',
'Backpropagation computes gradients during neural-network training.',
'A vector database can retrieve semantically related passages for a RAG pipeline.'
]
pairs = [(query, document) for document in documents]
scores = model.predict(pairs)
ranked = sorted(zip(scores, documents), reverse=True)
for score, document in ranked:
print(score, document)Mixedbread also provides the mxbai-rerank package, whose rank() method can return top-K documents directly.
MXBAI Reranker Large V2 reported benchmarks
Scores published by the model authors. A starting point, not a guarantee on your own data.
- BEIR Avg57.49
- Multilingual29.79
- Chinese84.16
- Code Search32.05
- NFCorpus A100 latency0.89 s/query
MXBAI Reranker Large V2 specifications
Architecture and serving details for MXBAI Reranker Large V2.
Architecture
- Type
- LLM-based cross-encoder reranker
- Backbone
- Qwen 2.5 family / Qwen2ForCausalLM architecture
- Parameters
- ~1.54B
- Layers
- 28
- Hidden size
- 1536
- Intermediate size
- 8960
- Attention
- 12 query heads / 2 KV heads
Context
- Recommended long context
- Up to 8K tokens
- Architectural maximum
- 32,768 positions
- Input
- Query + candidate document
- Output
- Relevance score
- Languages
- 100+ / 109 in Hugging Face metadata
- Code retrieval
- Supported
Training
- Stage 1
- GRPO
- Stage 2
- Contrastive learning
- Stage 3
- Preference learning
Serving
- Sentence Transformers
- CrossEncoder supported
- mxbai-rerank
- Native package
- Flash Attention 2
- Supported by mxbai-rerank
- Checkpoint config dtype
- BF16
- HexGrid model-specific API
- Not independently verified
MXBAI Reranker Large V2 frequently asked questions
Common questions about deploying MXBAI Reranker Large V2 on HexGrid Cloud.
Is MXBAI Reranker Large V2 a 1.5B or 2B model?
Mixedbread names it ProRank-1.5B and describes it as a 1.5B model. Its published Qwen-family configuration corresponds to roughly 1.54B parameters. Hugging Face may display the model size rounded more coarsely as 2B.
Does it support 32K context?
The checkpoint configuration has 32,768 maximum positions. Mixedbread describes normal V2 long-context support as up to 8K tokens and calls the models 32K-compatible. Their reranking library defaults to 8192 and can be increased up to the model limit.
How many languages does it support?
Mixedbread states more than 100 languages; Hugging Face metadata currently identifies 109 languages.
How fast is it?
Mixedbread measured an average 0.89 seconds per query on NFCorpus using an NVIDIA A100 80GB. That result is specific to their benchmark configuration and should not be presented as universal latency.
How was the V2 model trained?
Mixedbread describes a three-stage process: GRPO, followed by contrastive learning, followed by preference learning.
Can I use it commercially?
The model is released under Apache 2.0, which permits commercial use, modification and redistribution subject to the licence terms.
Deploy MXBAI Reranker Large V2 today
$0.20 per hour on 1× RTX A4000, billed per minute, never shared.