Zerank 2 Reranker
zeroentropy/zerank-2-reranker
published Nov 2025 · updated May 2026
Zerank 2 Reranker is a 4B parameter rerank model that scores query-document pairs to improve retrieval accuracy, outperforming proprietary rerankers across multiple domains.
specs
| Task | Reranking |
| Architecture | Cross-Encoder based on Qwen3-4B |
| Parameters | 4B |
| License | CC-BY-NC-4.0 |
about this model
zeroentropy/zerank-2-reranker is a 4B-parameter reranker model that scores query-document relevance using an Elo-inspired training methodology (zELO). It is built on Qwen3-4B, supports a context length of 32,768 tokens, and is designed to improve the accuracy of retrieval pipelines by re-ranking top candidates from any first-stage retriever (e.g., BM25, embedding-based search, or hybrid retrieval).
Key strengths
- Outperforms closed-source proprietary rerankers, including Cohere Rerank 3.5 and Gemini 2.5 Flash, across domains such as web, conversational, STEM, code, legal, biomedical, and finance.
- Multilingual instruction-following capability (as described in the official blog announcement).
- Trained using the zELO method on 112,000 queries with 100 documents per query (over 5 million pairs) in less than 10,000 H100-hours.
Benchmark performance (NDCG@10)
All rerankers evaluated with OpenAI text-embedding-3-small as the first-stage retriever (Top 100 candidates).
| Domain | OpenAI embeddings | zerank-2 | zerank-1 | Gemini 2.5 Flash (Listwise) | Cohere Rerank 3.5 |
|---|---|---|---|---|---|
| Web | 0.3819 | 0.6346 | 0.6069 |
best for
- ·Improving search result accuracy by reranking top candidates from a first-stage retriever
- ·Domain-specific retrieval in finance, legal, code, STEM, and biomedical applications
FAQ
The model supports a context length of 32,768 tokens (32k).
It outperforms Cohere rerank-3.5 and Gemini 2.5 Flash on average NDCG@10 across web, conversational, STEM, code, legal, biomedical, and finance domains.
It is released under the CC-BY-NC-4.0 non-commercial license. A commercial license is available by contacting ZeroEntropy.
Use the gigarouter OpenAI-compatible endpoint with your API key to send query-document pairs for scoring.
The model expects query-document text pairs. It can be used via Sentence Transformers with model.predict() or model.rank(), or via the gigarouter API.
We're benchmarking and onboarding Zerank 2 Reranker 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.