What it is
zerank-2 is a reranker from ZeroEntropy, announced on 18 November 2025. Notion acquired ZeroEntropy on 24 July 2026; the same day the company released all its models under Apache 2.0 and said its hosted products would be supported until 4 September 2026. That date has passed, and ZeroEntropy's API now answers rerank requests with "ZeroEntropy inference is unavailable", so the model is in practice self-host only.
A reranker rescores the shortlist a search system has already retrieved by reading the query and each document together, which is slower than comparing vectors but more precise. zerank-2 follows natural-language instructions about what counts as relevant and is trained so its scores are calibrated: ZeroEntropy says a score of 0.8 means roughly 80% relevance. Training uses the company's zELO method, which turns pairwise preferences into Elo-style ratings.
It has 4 billion parameters, about 8 GB of BF16 weights, is fine-tuned from Qwen3-4B and reads up to 32,768 tokens; ZeroEntropy claims more than 100 languages. It loads as a sentence-transformers CrossEncoder. A May 2026 change made predict() return raw logits rather than 0-to-1 scores; the card gives the conversion, (scores / 5).sigmoid().
Our review
zerank-2 earns its place through calibration. Most rerankers return scores that only make sense as an order; this one is trained so the number itself means something, which lets a pipeline drop weak passages at a fixed threshold, or use the model as a yes-or-no relevance check before context reaches a language model.
On its model card, reranking the top 100 results from OpenAI's text-embedding-3-small, it averages 0.6714 NDCG@10 across seven domains, against 0.5847 for Cohere rerank-3.5 and 0.5999 for Gemini 2.5 Flash. It loses only finance, where Gemini scores 0.7694 to its 0.7600. The numbers are ZeroEntropy's, and the rivals are a generation old: Cohere Rerank 4 and Voyage rerank-3 are not in the table. Notion's acquisition post adds that the reranker cut its reranking latency by 85% at constant answer quality, a customer's word rather than a benchmark.
The circumstances have moved more than the model. ZeroEntropy's team now runs Notion's model research, the hosted API has stopped serving, and the newest post on its news page is the July acquisition notice, with no sign of a zerank-3. Apache 2.0 removes the licensing risk and the card still lists AWS Marketplace, but fixes and new versions now depend on whoever forks it.
Two migration details catch people out. The May 2026 switch to raw logits silently breaks any threshold tuned on the old 0-to-1 scores unless the sigmoid conversion is applied, and the card's examples load "zeroentropy/zerank-2" while the repository is named zerank-2-reranker. Neither is hard to fix, but both fail without an error message, which is exactly how a retrieval pipeline degrades unnoticed.
Where it earns its place
A RAG engineer at a legal-research start-up
Cut irrelevant passages before they reach the language model
Rerank the top 100 hits and keep only those above a calibrated cut-off after the sigmoid conversion, so each answer carries fewer input tokens. Legal is one of the domains with the widest lead on the card: 0.6644 against 0.5257 for Cohere rerank-3.5.
A former ZeroEntropy API customer
Keep search running after the hosted API stopped serving
Follow ZeroEntropy's July migration guide, serve the Apache 2.0 weights on one GPU behind your own endpoint, and re-tune thresholds against the raw-logit output. The cost moves from $0.025 per million tokens to whatever an always-on GPU instance costs.
A search lead at a 40-person SaaS company
Choose between a self-hosted reranker and a managed API
zerank-2 is free to run but leaves upgrades, fixes and uptime to the team; Cohere Rerank 4 or Voyage rerank-3 bill per query but come with a vendor. The deciding factor is usually whether someone already runs GPU inference in production.
Before you commit
- Setup
- pip install sentence-transformers, load the model as a CrossEncoder and call predict() or rank(); trust_remote_code is no longer needed. Apply (scores / 5).sigmoid() to recover calibrated 0-to-1 scores.
- Integrations
- Hugging Face weights with 15 community quantisations, AWS Marketplace according to the model card, and any server that runs sentence-transformers or Qwen3-based models. ZeroEntropy's hosted /models/rerank endpoint no longer serves requests.
- Data and privacy
- Apache 2.0 allows commercial use, modification and redistribution with no revenue limits. Self-hosting keeps queries and documents inside your own infrastructure; there is no longer a vendor API to send data to.
- Cost at scale
- The weights are free; the cost is a GPU holding about 8 GB of BF16 weights plus batch memory, which a single 24 GB data-centre card covers. Latency and throughput figures are not published, so capacity has to be measured on your own query mix.
- Limits
- 32,768-token context per query-document pair; text only. No published latency figures, and no maintainer roadmap since the Notion acquisition.
What's good
- Calibrated scores usable as a fixed cut-off
- Apache 2.0 weights since July 2026
- 32K-token context and natural-language relevance instructions
- Beats Cohere rerank-3.5 in six of seven domains on its card
What's not
- Hosted API no longer serving; self-host only
- No vendor roadmap after the Notion acquisition
- Benchmarks are ZeroEntropy's own, against older rivals
- May 2026 switch to raw logits breaks old thresholds
Best for
Teams already running GPU inference who want a free reranker with relevance scores they can use as a cut-off.
Alternatives
zerank-2: common questions
- Is zerank-2 worth it?
- A sharp open reranker whose calibrated scores are worth having, now with nobody behind it but your own GPU. We rate it 3.9 out of 5.
- How much does zerank-2 cost?
- Free under Apache 2.0. ZeroEntropy's hosted API ($0.025 per 1M tokens at launch) was supported only until 4 September 2026 and now returns an unavailable error.
- What are the downsides of zerank-2?
- Hosted API no longer serving; self-host only; no vendor roadmap after the Notion acquisition; benchmarks are ZeroEntropy's own, against older rivals; may 2026 switch to raw logits breaks old thresholds.
- Who is zerank-2 best for?
- Teams already running GPU inference who want a free reranker with relevance scores they can use as a cut-off.
- How much does zerank-2 cost for a team or in production?
- The weights are free; the cost is a GPU holding about 8 GB of BF16 weights plus batch memory, which a single 24 GB data-centre card covers. Latency and throughput figures are not published, so capacity has to be measured on your own query mix.
- Is zerank-2 safe for company data?
- Apache 2.0 allows commercial use, modification and redistribution with no revenue limits. Self-hosting keeps queries and documents inside your own infrastructure; there is no longer a vendor API to send data to.
- What are the best zerank-2 alternatives?
- The closest alternatives we have reviewed are Cohere Rerank 4, Voyage rerank-3, Jina Reranker v3.5 and mxbai-rerank v2.
Answers checked Oct 22, 2026.