GTE Multilingual Reranker Base

The gte-multilingual-reranker-base is a cross-encoder reranking model from the GTE family, designed to reorder document lists by relevance to a given query. It is built on an encoder-only transformer architecture (306M parameters) and supports input sequences up to 8,192 tokens, enabling long-context retrieval tasks. The model covers over 70 languages.

Performance and Efficiency

The model achieves state-of-the-art results on multilingual retrieval benchmarks among rerankers of comparable size. According to the accompanying paper (EMNLP 2024 Industry Track), the reranker matches the performance of the larger BGE-M3 models and surpasses them on long-context retrieval benchmarks. Its encoder-only design provides an approximately 10x inference speed advantage over decoder-only LLM-based rerankers (e.g., gte-qwen2-1.5b-instruct) and requires lower hardware resources.

Architecture and Training

The underlying text encoder is pre-trained with a native 8,192-token context (compared to 512 tokens for previous multilingual encoders like XLM-R) and enhanced with Rotary Position Embedding (RoPE) and unpadding optimization. The reranker is trained via contrastive learning on a hybrid of text representation and cross-encoder objectives.

Benchmark Results

Evaluation on multiple text retrieval datasets demonstrates the model's effectiveness. Detailed experimental results are available in the paper.

This model is hosted as a managed, OpenAI-compatible API on gigarouter, requiring no local infrastructure or model loading.