AI Potluck
Contribute
Back to Gap Map Model components / Embeddings & retrieval models

GTE Reranker

Alibaba Cloud
open weights / Overall score: 3.3

The reranking half of Alibaba's GTE family: two Apache-2.0 cross-encoders that score query-document pairs for a retrieval pipeline rather than emit vectors. gte-reranker-modernbert-base is a 149M English model on a ModernBERT base reporting BEIR 56.19, LoCo 90.68 and CoIR 79.99 at 8,192 tokens, and gte-multilingual-reranker-base is a 306M encoder-only reranker covering more than 70 languages at the same context, which the mGTE paper positions as roughly ten times faster than the decoder-based alternatives of its generation.

Added 2026-09-11 for the embeddings_retrieval roster, splitting the GTE cross-encoders out from the sibling gte product, which covers the embedders. Scope is the two reranker checkpoints in the Alibaba-NLP account and nothing else; the gte-rerank service on Alibaba Cloud is a separate hosted product the cards note is not identical to the open weights. Verified 2026-09-11 via the model cards and the Hub API.

Openness

3 high confidence
3.0
weights
open(Apache-2.0 ungated safetensors for both cross-encoders, with ONNX builds beside the ModernBERT one)
data
closed(both cards defer the training mixture to the mGTE paper and release no corpus)
code
partial(inference usage through transformers, TEI and Infinity ships on the cards
license
Apache-2.0(OSI)

The same resolution as the gte embedders, on the same evidence: an OSI license over ungated weights, with the corpus and the training run left in the mGTE paper. No reranker SKU carries a more restrictive license, so nothing pulls the line down.

Adoption

4 high confidence
4.0

3,221,049 downloads in the trailing 30 days across the two reranker checkpoints on the Hugging Face universe sweep - gte-reranker-modernbert-base 2,728,127 and gte-multilingual-reranker-base 492,922. The ModernBERT cross-encoder alone clears the 1M floor. Live Hub API figures a few days later read 2,638,520 and 372,525 and do not move the band. The gte embedders are the sibling gte product and are not counted here.

Capability

3 high confidence
3.0

Rung 3, solid workhorse. Heavily used and credibly measured, but a band below the reranking frontier on the same instrument: 56.19 BEIR nDCG@10 against jina-reranker-v3 at 61.94, and the ModernBERT cross-encoder that carries the downloads is English only where the frontier rerankers are multilingual. The multilingual sibling reaches 70+ languages but is the 2024 generation. Same placement and the same peer as bge-reranker, which sits one below jina-reranker on the same reading. The gte embedders are scored separately at 4; this edge is about the reranking half only.

Verified 2026-09-11