GTE rerankers

Open weights · Tongyi Lab, Alibaba · Updated 16 Oct 2026

Alibaba’s two GTE rerankers are small encoder cross-encoders, not LLM-based ones: gte-reranker-modernbert-base (149M, English) and gte-multilingual-reranker-base (306M, 70+ languages). Both take 8,192 tokens and are Apache 2.0. They are the size you can serve on a CPU or a small GPU, and the English model even runs in the browser.

The two models

Alibaba Cloud’s hosted gte-rerank APIs are related models, not these weights; see reranking in mainland China for the hosted ones.

Scores, and whose they are

Alibaba’s card for the ModernBERT model doesn’t say which BEIR datasets or which first stage it used. For the multilingual model Alibaba publish results only as a chart (on the card and in the mGTE paper), so the numbers here are Qwen’s, who ran it as a baseline for Qwen3-Reranker, reranking the top 100 from Qwen3-Embedding-0.6B. The two models’ rows aren’t comparable with each other.

Next to Alibaba’s newer Qwen3-Reranker, the multilingual model is the cheap end of the scale: on the same Qwen runs the 0.6B Qwen3-Reranker scores 65.80* on MTEB-R and 71.31 on CMTEB-R, for twice the parameters and an LLM-style serving setup.

Quick start

sentence-transformers

pip install sentence-transformers

from sentence_transformers import CrossEncoder

model = CrossEncoder(
    "Alibaba-NLP/gte-reranker-modernbert-base",
    automodel_args={"torch_dtype": "auto"},
)
pairs = [
    ["what is the capital of China?", "Beijing"],
    ["how to implement quick sort in python?", "Introduction of quick sort"],
]
scores = model.predict(pairs)   # 0–1: sentence-transformers applies an activation

The multilingual model’s card loads it with plain transformers and trust_remote_code=True, and recommends installing xformers and enabling unpadding for speed.

As a service

docker run -p 8080:80 -v $PWD/data:/data \
  ghcr.io/huggingface/text-embeddings-inference:cpu-1.7 \
  --model-id Alibaba-NLP/gte-reranker-modernbert-base

curl localhost:8080/rerank -H "Content-Type: application/json" \
  -d '{"query": "What is the capital of China?", "texts": ["Beijing"]}'

Hugging Face’s Text Embeddings Inference serves both models behind a /rerank route, on CPU or GPU (drop cpu- from the image tag and add --gpus all). The English model also runs in the browser through transformers.js, in fp32, fp16, q8, q4 or q4f16.

When to pick GTE

• Yes: CPU serving, or a GPU budget too small for an LLM-based reranker.
• Yes: long passages: both take 8,192 tokens, where many small cross-encoders stop at 512.
• Yes: English code or long-document search, the two areas Alibaba report for the ModernBERT model.
• Maybe not: you need instructions, or the best multilingual quality and have a GPU: Qwen3-Reranker takes instructions and scores higher on Qwen’s runs.

Pros and cons

Other models