GTE Reranker ModernBERT Base
Alibaba's 149M-parameter open reranker on ModernBERT: 8K context, Apache 2.0, CPU-friendly with ONNX exports.
About
gte-reranker-modernbert-base is Alibaba Tongyi Lab's open-weight text reranker, released January 21, 2025 as part of the gte-modernbert series. It's a 149M-parameter cross-encoder built on ModernBERT-base with an 8,192-token input window — 16x the context of the classic MiniLM rerankers — and it ships under Apache 2.0 with pre-exported ONNX variants, so it runs comfortably on CPU with no PyTorch at runtime. The model card reports BEIR 56.19, LoCo 90.68, and CoIR 79.99.
It has become a quiet default in self-hosted RAG setups: several open-source projects ship it as their bundled local reranker because it's small, fast, permissively licensed, and good enough to meaningfully lift retrieval quality over first-stage vector search. The tradeoff is that it's English-primary and hasn't seen vendor updates since release, so for multilingual or very long-document reranking you'd look at bigger options.
Specs
Context window
Pricing
Free quota
API access
Weights
Released
Benchmarks
- BEIR 56.19, LoCo 90.68, CoIR 79.99 (Alibaba-reported on the Hugging Face model card)
- Third-party analyses note it matches ~1.2B-parameter rerankers on Hit@1 at 8x smaller size (community-reported, not vendor-verified)
Provider-reported unless a source says otherwise. Benchmarks are marketing-adjacent by nature — treat them as directional.
Use cases
Trust score — 7/10
Scored 7/10 (metrics tier, metrics-v1). Evidence: HF model page released 2025-01-21 with Apache 2.0 license metadata; architecture (ModernBERT-base, 149M, 8192 ctx) corroborated by independent RAG project docs and deployment guides; 8 pre-exported ONNX variants confirmed on the Hub. Open weights. Strengths: permissive license, genuinely CPU-deployable, adopted as default local reranker by multiple open-source RAG projects. Docked for: (1) benchmark figures are vendor-reported on the model card with no independent replication found; (2) English-primary, no vendor updates since the January 2025 release. License: Apache-2.0.
How the trust score is built
Full methodology →- Adoption · stars/downloads, log-scaled vs peers
- 7.0/10
- Maintenance · recency, releases, issue load
- 6.0/10
- Openness · license + docs
- 9.0/10
- Manual review · sanity read + red-flag check
- 7.0/10
- · Released 2025-01-21 by Alibaba Tongyi Lab; 149M params, 8192-token context, Apache 2.0
- · BEIR 56.19 / LoCo 90.68 / CoIR 79.99 (Alibaba-reported on HF model card)
- · Ships 8 pre-exported ONNX variants; adopted as default local reranker by several open-source RAG projects
- · English-primary; no vendor updates since release