MXBAI Reranker Base V2
mixedbread-ai/mxbai-rerank-base-v2
published Mar 2025 · updated Apr 2026
MXBAI Reranker Base V2 is a rerank model that uses reinforcement learning to improve document reranking quality while maintaining computational efficiency.
specs
| Task | Reranking |
| Architecture | Transformer small language model (0.5B parameters) |
| Parameters | 0.5B |
| License | Apache 2.0 |
about this model
mxbai-rerank-base-v2 is a reranking model for information retrieval and retrieval-augmented generation, optimized for high efficiency and strong benchmark performance. Built on a 0.5-billion-parameter small language model (SLM), it uses a three-stage training pipeline—GRPO (Guided Reinforcement Prompt Optimization), contrastive learning, and preference learning—to overcome limitations of small models in representation expressiveness and prompt understanding, as described in the ProRank paper (accepted at ACL 2026 Findings).
Key Capabilities
- Supports 100+ languages and up to 8,000 tokens context (with 32,000-token compatibility).
- Accepts custom reranking instructions for domain-specific tasks.
- Licensed under Apache 2.0.
Benchmark Performance
Evaluated on standard information retrieval benchmarks, base-v2 achieves strong results with low latency:
| Benchmark | Base-v2 (0.5B) | Large-v2 (1.5B) |
|---|---|---|
| BEIR Average | 55.57 | 57.49 |
| Multilingual | 28.56 | 29.79 |
| Chinese | 83.70 | 84.16 |
| Code Search | 31.73 | 32.05 |
Latency on an A100 GPU is 0.67 seconds for base-v2, compared to 0.89 seconds for large-v2. The v2 family significantly improves over the legacy v1 models (e.g., large-v1 BEIR average 49.32, latency 2.24 seconds) and is reported to be 8× faster than comparable SLM rerankers. The underlying ProRank approach demonstrates that properly trained SLMs can surpass larger reranking models on the BEIR benchmark while maintaining computational efficiency.
best for
- ·Document retrieval reranking
- ·Retrieval-Augmented Generation (RAG) reranking
- ·Multilingual search reranking
FAQ
Supports up to 8,000 tokens (32k-compatible).
It achieves a BEIR average of 55.57 while being 8x faster than comparable models.
It uses ProRank, a two-stage approach with reinforcement learning and fine-grained score learning.
Yes, it supports 100+ languages.
Use the gigarouter OpenAI-compatible endpoint with an API key.
We're benchmarking and onboarding MXBAI Reranker Base V2 as a hosted, OpenAI-compatible API. Sign in for free credit and be ready when it lands, or tell us you want it and we'll prioritize it.