Crispy reranking models from Mixedbread. State-of-the-art models for search relevance, powered by reinforcement learning.

• State-of-the-art performance - Outperforms leading open and closed-source rerankers on major benchmarks
• 100+ languages - Strong multilingual support out of the box
• Long context - Handle up to 8k tokens (32k-compatible)
• Code & SQL - Excellent at ranking code snippets and technical content
• Function Call Ranking - Supports reranking of function calls for multi-tool agents
• Fast inference - 8x faster than comparable models
• Easy integration - Drop-in improvement for existing search systems
• Open source - Apache 2.0-licensed, easy to customize
• Managed API - For production use with additional features. We support embeddings, reranking, and an end-to-end multi-modal retrieval solution.

pip install -U mxbai-rerank

from mxbai_rerank import MxbaiRerankV2

# Initialize the reranker
reranker = MxbaiRerankV2("mixedbread-ai/mxbai-rerank-base-v2")  # or large-v2

# Example query and documents
query = "Who wrote 'To Kill a Mockingbird'?"
documents = [
    "'To Kill a Mockingbird' is a novel by Harper Lee published in 1960.",
    "The novel 'Moby-Dick' was written by Herman Melville.",
    "Harper Lee was born in 1926 in Monroeville, Alabama."
]

results = reranker.rank(query=query, documents=documents)

print(results)

We offer multiple model variants. For more details, see our mxbai-rerank-v2 technical blog post.

• mxbai-rerank-base-v2 (0.5B) - Best balance of speed and accuracy
• mxbai-rerank-large-v2 (1.5B) - Highest accuracy, still with excellent speed

For more details, see our mxbai-rerank-v1 technical blog post.

• mxbai-rerank-xsmall-v1 (0.1B) - Fastest inference, lower accuracy
• mxbai-rerank-base-v1 (0.2B) - Smaller, faster model
• mxbai-rerank-large-v1 (1.5B) - Large model with highest accuracy

*Latency measured on A100 GPU

The v2 models automatically use Flash Attention 2 when available for faster inference:

pip install flash-attn --no-build-isolation

reranker = MxbaiRerankV2(
    "mixedbread-ai/mxbai-rerank-base-v2",
    max_length=8192  # Default, can be adjusted up to model limits (32k for v2 models)
)

results = reranker.rank(query=query, documents=documents, instruction="Figure out the best code snippet for the user query.")

For managed API access with additional features, such as object reranking and instructions:

from mixedbread import Mixedbread

mxbai = Mixedbread(api_key="YOUR_API_KEY")

results = mxbai.rerank(
    model="mixedbread-ai/mxbai-rerank-large-v2",
    query="your query",
    input=["doc1", "doc2", "doc3"]
)

The models were trained using a three-step process:

1. GRPO (Guided Reinforcement Prompt Optimization)
2. Contrastive Learning
3. Preference Learning

For more details, check our technical blog post or preprint paper.

Paper following soon.

If you use this work, please cite:

@article{li2025prorank,
  title={ProRank: Prompt Warmup via Reinforcement Learning for Small Language Models Reranking},
  author={Li, Xianming and Shakir, Aamir and Huang, Rui and Lipp, Julius and Li, Jing},
  journal={arXiv preprint arXiv:2506.03487},
  year={2025}
}

This project is licensed under the Apache License 2.0 - see the LICENSE file for details.

Contributions are welcome! Please feel free to submit a pull request or report an issue on GitHub.

• Discord Community
• X/Twitter
• LinkedIn