Model description

mxbai-rerank-base-v2 is part of the second-generation Mixedbread rerank model family, a set of state-of-the-art reranking models that are fully open-source under the Apache 2.0 license. This 0.5B-parameter model provides an excellent balance of size, speed, and performance, making it ideal for most reranking applications.

The v2 models represent a significant advancement over the first generation, featuring:

• Reinforcement learning training - Using GRPO (Guided Reinforcement Prompt Optimization) for enhanced performance
• Multilingual capabilities - Supporting 100+ languages for global applications
• Extended context handling - Processing up to 8K tokens (32K-compatible) for comprehensive document analysis
• Complex query reasoning - Improved understanding of nuanced search intent
• Versatile applications - Excelling at code search, SQL ranking, and function-call reranking for multi-tool agents

On benchmarks, mxbai-rerank-base-v2 achieves exceptional results with an NDCG@10 score of 55.57 on BEIR average, 28.56 on Mr.TyDi multilingual datasets, 83.70 on Chinese datasets, and 31.73 on code search tasks. It delivers this performance with impressive speed, processing queries up to 8x faster than comparable solutions.

When used in combination with a keyword-based search engine such as Elasticsearch, OpenSearch, or Solr, the rerank model can be added to the end of an existing search workflow. This allows users to incorporate semantic relevance into their keyword-based search system without changing the existing infrastructure - an easy, low-complexity method of improving search results with just one line of code.

Key Advantages

• Exceptional performance-to-size ratio - Offering near state-of-the-art performance in a compact 0.5B model
• Fast processing - Delivers results with minimal latency (0.67s per query on NFC dataset with A100 GPU)
• Reinforcement learning enhanced - Trained with advanced GRPO techniques for improved relevance scoring
• Versatile applications - Beyond document search, excels at code snippets, SQL, LLM tool selection, and e-commerce ranking

Limitations

• Resource Requirements: While more efficient than larger models, still requires appropriate GPU resources for optimal performance.
• Sequence Truncation: Documents exceeding the 32K token limit will be truncated, which may result in information loss. Please note that max sequence length is for the query and document combined. It means that len(query) + len(document) should not be longer than 32K tokens.

Examples