GTE Reranker ModernBERT Base
Alibaba-NLP/gte-reranker-modernbert-base
published Jan 2025 · updated Jul 2025
GTE Reranker ModernBERT Base is a rerank model that re-ranks documents based on relevance to a query, built on the ModernBERT encoder.
specs
| Task | Text Reranker |
| Architecture | Cross-encoder (ModernBERT encoder) |
| Parameters | 149M |
| License | Apache-2.0 |
| Max Input Length | 8192 tokens |
| Primary Language | English |
about this model
Alibaba-NLP/gte-reranker-modernbert-base is a text reranker model that scores query-document pairs to produce relevance rankings, built on the modernBERT encoder with 149M parameters and a native 8192-token context window.
Key Strengths
The model achieves strong results across retrieval benchmarks, particularly for long-document and code search tasks. It uses RoPE and unpadding for efficient long-context processing and is trained with multi-stage contrastive learning following the mGTE methodology (EMNLP 2024).
Benchmark Performance
| Benchmark | Score |
|---|---|
| BEIR (average over 15 datasets) | 56.19 |
| LoCO (long-document retrieval, average over 5 tasks) | 90.68 |
| CoIR (code retrieval, average over 20 tasks) | 79.99 |
Notable per-task results include LoCO QsmsumRetrieval: 70.86, SummScreenRetrieval: 94.06, QasperAbastractRetrieval: 99.73; CoIR CodeSearchNet-ccr-go: 96.43, CodeSearchNet-ccr-java: 96.88.
Technical Details
- Model size: 149M parameters
- Maximum input length: 8192 tokens
- Primary language: English
- License: Apache 2.0
- Developed by Tongyi Lab, Alibaba Group
The model is referenced in the paper mGTE: Generalized Long-Context Text Representation and Reranking Models for Multilingual Text Retrieval and builds on the GTE training scheme using modernBERT as the base.
best for
- ·Reranking search results for long-document retrieval
- ·Code retrieval reranking
- ·Improving relevance for enterprise search and QA systems
FAQ
It is best for reranking documents to improve search relevance, especially for long documents and code retrieval tasks.
It has 149M parameters, uses Flash Attention 2 for efficiency, and supports up to 8192 tokens input.
Apache-2.0.
Use the gigarouter OpenAI-compatible endpoint with an API key and send queries and documents in the rerank format.
It expects pairs of query and document text, and returns relevance scores.
We're benchmarking and onboarding GTE Reranker ModernBERT Base 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.