Back to Gap Map Model components / Embeddings & retrieval models

zerank

ZeroEntropy
open weights / Overall score: 3.7

ZeroEntropy's reranker line. zerank-2 is a 4B Qwen3-derived cross-encoder with a 32K context, trained through a multi-stage pipeline that models query-document relevance as adjusted Elo ratings; its card reports an average 0.6714 NDCG@10 across seven domains against 0.5847 for Cohere rerank-3.5. The earlier zerank-1 and zerank-1-small ship alongside it. Notion acquired ZeroEntropy in July 2026 and the weights were released under Apache 2.0.

The hosted ZeroEntropy API was sunset on 2026-09-04; the weights themselves are unaffected and remain on the Hub, so this entry carries no end-of-life mark.

Openness

3 high confidence
3.0
weights
open(Apache-2.0 ungated safetensors for zerank-1, zerank-1-small and zerank-2)
data
closed(the Elo-based relevance pipeline is described in arXiv 2509.12541
code
closed(no training or model repository
license
Apache-2.0(OSI)

Both zerank-2 and the earlier zerank-1 line are released under Apache-2.0, confirmed in the card and in ZeroEntropy's announcement of the Notion acquisition, which states that all its models are now open source under Apache-2.0. Neither the training data nor any training or model repository is released, so the release stays open weights rather than open source.

Adoption

3 high confidence
3.0

Adoption is measured as Hugging Face downloads across the zerank line, almost all of it for zerank-2, with much smaller shares for zerank-1-small and zerank-1.

Capability

4 medium confidence
4.0

Zerank-2 is a current 4B model that beats the named proprietary rerankers on ZeroEntropy's own seven-domain suite, putting it at the frontier alongside jina-reranker, the other current-generation line here. Confidence stays medium because that evaluation is a vendor-built suite with no public leaderboard behind it, unlike BEIR or MMTEB.

  • https://huggingface.co/zeroentropy/zerank-2-reranker/raw/main/README.md recorded 2026-09-11

    Evaluations table gives zerank-2 an average NDCG@10 of 0.6714 across seven domains against 0.6456 for zerank-1, 0.5999 for Gemini 2.5 Flash listwise, 0.5847 for Cohere rerank-3.5 and 0.4509 for the OpenAI embedding retriever alone; 4B parameters, 32,768-token context.

Verified 2026-09-11