Notion acquires ZeroEntropy, the retrieval-model startup behind its search reranker

What ZeroEntropy built

ZeroEntropy was founded in 2024 and went through Y Combinator's Winter 2025 batch. It built retrieval infrastructure for AI applications, including rerankers, embedding models and custom models for context compression and query rewriting. Its products were available through an API, inside a customer's virtual private cloud, or as on-premises or dedicated deployments.

Its best-known models were the zerank family of rerankers, including zerank-2, and the zembed-1 embedding model. The team also published a research paper on a training method it called zELO.

Notion described the team's work as high-efficiency training and inference for task-specific models, including custom kernels, its own inference stack and recipes for specialising models. Rather than relying mainly on large general-purpose models, ZeroEntropy used smaller models for individual steps such as reranking and routing. In July 2025 it was used by more than 10 early-stage AI companies in healthcare, law, customer support and sales.

From supplier to acquisition

Notion first worked with ZeroEntropy as a customer. Notion AI's unified search ran on the zerank-2 reranker, and Notion said the model made that search up to 30% faster and cut latency on the reranking step by 85%, with answer quality unchanged and lower inference costs.

Notion said the team shared its aim of bringing useful AI to businesses of every size, not only to those spending millions of dollars a month on model tokens, and that fast, inexpensive, task-specific models help make that possible.

The deal is classed here as an acquihire: the team moves to Notion, the models are released publicly, and ZeroEntropy's commercial products are being wound down.

What happens to customers

ZeroEntropy published a migration guide for customers of its API. It said the zerank-2, zerank-1 and zerank-1-small rerankers and the zembed-1 embedding model were available as Apache 2.0 open weights on Hugging Face, so customers could host them themselves and get identical scores.

Customers that prefer a managed service were pointed to other providers' reranking and embedding models, with the warning that score scales differ, so ranking thresholds would need to be re-tuned and document collections re-embedded. Users of ZeroEntropy's hosted search product were advised to rebuild on a search engine or vector database they run themselves, and to export text the service had already extracted from their files.

The founders and the funding

Houir Alami was born and raised in Morocco and left at 17 to study engineering in France. She holds master's degrees in applied mathematics from École Polytechnique and the University of California, Berkeley. Her co-founder, Nicholas Pipitone, served as chief technology officer. Notion said the wider team had backgrounds in AI research, quantitative finance and physics.

In July 2025 ZeroEntropy raised a $4.2 million seed round led by Initialized Capital, with Y Combinator, Transpose Platform, 22 Ventures and a16z Scout participating, along with individual investors from OpenAI, Hugging Face and Front. The company's own announcement put the round at $4.1 million. No later funding was reported.

Software companies build their own models

The acquisition gives Notion an in-house team for training specialised models instead of relying only on outside model providers for every step of its AI features. Reranking, the step that decides which retrieved passages are most relevant to a question, was the part of Notion's search where ZeroEntropy's model had already cut latency and cost.

Notion did not say which models the new research team would build first or whether Notion AI would continue to use the open-sourced zerank models.