ZeroEntropy: Revenue, sale and how it was built
ZeroEntropy (Reranker and embedding models sold through an API for AI search, United States): Acquired by Notion with its whole team, July 2026. Evidence grade E: no revenue figure was checked. Team: 2 co-founders in 2024; Y Combinator lists a team of 6 (read November 2026); all joined Notion in July 2026. Launched May 2025; sold 14 months after launch.Added 7 November 2026
- Outcome
- Acquired by Notion with its whole team, July 2026
- Evidence
- E no revenue figure was checked
- Launch
- May 2025
- Status
- Sold, 14 months after launch
- Team
- 2 co-founders in 2024; Y Combinator lists a team of 6 (read November 2026); all joined Notion in July 2026
- Outside funding
- Pre-seed from Entrepreneur First, Y Combinator Winter 2025, and a US$4.2M seed led by Initialized Capital (July 2025); about US$4.4M in total, the company said
- Main channel
- Not stated
What is not verified
No sale price, revenue figure or count of paying customers was found. We read Notion's announcement, ZeroEntropy's announcement and migration guide, its pricing, blog and about pages, and its Y Combinator page; none gives an amount. The team size of 6 comes from Y Combinator's directory, with no date. A reader could look for a post by the lead investor, Initialized Capital, or a later interview with the founders. This story is grade E: the business is real and its product, price and customers are described from public pages, but no revenue figure was checked. Aiolum shows it as an idea worth studying, not as proof that it makes money.
ZeroEntropy sold models for retrieval, the step in which a program finds the few documents that an AI model should read before it answers; when those documents are then given to a language model, the method is called retrieval-augmented generation (RAG). Its main products were a reranker, a model that takes a question and a list of candidate documents and gives each document a relevance score so that the best ones come first, and an embedding model, which turns a text into a list of numbers so that texts with similar meaning get similar numbers. Developers called them through an API and paid per million tokens (pieces of words) processed, for example US$0.025 per million tokens for the reranker, as the July 2025 launch post and the pricing page state.
Ghita Houir Alami (CEO), who holds two master's degrees in applied mathematics, and Nicholas Pipitone (CTO) founded it in 2024, with a pre-seed investment from Entrepreneur First, and went through Y Combinator's Winter 2025 batch; the company is based in San Francisco. In July 2025 the company said it had raised a US$4.2 million seed round led by Initialized Capital. On 24 July 2026 Notion announced that it had acquired ZeroEntropy, that the whole team would join, and that Houir Alami would lead a new Model Research team.
Notion did not keep the product. A sale in which the buyer mainly wants the team and the product stops is called an acqui-hire: ZeroEntropy closed new signups the same day, said its products would be shut down after 4 September 2026, and moved its models to the Apache 2.0 license, which lets anyone use them, also commercially. No sale price, revenue or count of paying customers was published, so this story has grade E: a real business with checked facts but no money figure.

How much did ZeroEntropy make, and what was it sold for?
Each row shows the evidence grade and the source numbers from the list at the end of this page. The grades are explained in How Aiolum verifies.
| Fact | Value | Grade | How we know | Sources |
|---|---|---|---|---|
| Outcome, 24 July 2026 | acquired by Notion; the whole team joined; price not disclosed | A | Notion's own announcement: "Today we're announcing that we've acquired ZeroEntropy" and "Their entire team will be joining Notion". ZeroEntropy's founder posted the same news the same day. Neither post states a price. | [1, 2] |
| What happened to the product | new signups closed; products supported until 4 September 2026, then shut down | C | ZeroEntropy's announcement: "All ZeroEntropy products remain fully supported until September 4th, 2026, after which they will be sunset. New signups are disabled." A migration guide sent customers to other vendors or to self-hosting. | [2, 3] |
| Models after the sale | all model weights under the Apache 2.0 license | C | The announcement and the migration guide. Before the sale, zerank-1 had a non-commercial license and zerank-1-small was Apache 2.0. | [2, 3, 6] |
| Founders | Ghita Houir Alami (CEO) and Nicholas Pipitone (CTO) | A | Named by Notion as "founder and CEO" and "her cofounder and CTO", and listed as the two founders on the Y Combinator page. | [1, 4] |
| Founded | 2024; based in San Francisco; Y Combinator Winter 2025 | C | Y Combinator company page, read 7 November 2026, which now shows the status "Acquired". Notion wrote that the team built its technology "in less than two years". | [4, 1] |
| Team size | 6 | C | Y Combinator company page, read 7 November 2026, with no date for the figure. Notion called it "This tiny team". | [4, 1] |
| Funding | US$4.2 million seed, July 2025; about US$4.4 million in total | C | Company's post of 9 July 2025: seed led by Initialized Capital with Y Combinator and angel investors; Entrepreneur First "pre-seeded the company". | [5] |
| First reranker price, July 2025 | US$0.025 per million tokens | C | Launch post of zerank-1, 10 July 2025: "half the cost of Cohere's rerank-3.5". The same price was on the pricing page for zerank-2 in November 2026. | [6, 7] |
| Notion's results with zerank-2 | unified search up to 30% faster; reranking latency down 85% | A | Notion's announcement. ZeroEntropy wrote that Notion AI's unified search had run on zerank-2 "for the past 2 months" and that Notion was "one of our top customers". | [1, 2] |
| Named customers | Notion, Assembled, Vera Health, Mem0, My AskAI, Equall | C | Case studies on the company blog from September 2025 to February 2026, and the acquisition post. No count of paying customers was published. | [10, 2] |
| Revenue and sale price | not published | C | Not found in Notion's post, ZeroEntropy's posts, its pricing page or the Y Combinator page. | [1, 2, 4, 7] |
| Time from the first dated product description to the sale | 14 months | calc | Calculated: May 2025 (the "ZeroEntropy v1" architecture post of 14 May 2025) to July 2026. | [8, 1] |
How to read this table: a grade A value is stated by the buyer, Notion, in its own announcement. A grade C value is a statement by ZeroEntropy on its own pages or by the company on its Y Combinator page. A "calc" value was computed on this page. No money amount for the sale or for revenue exists in the sources.
| Plan | Per month, billed yearly (US$) | Per month, billed monthly (US$) | Limit |
|---|---|---|---|
| zerank-2 reranker | 0.025 per million tokens | rate limit 2,500,000 UTF-8 bytes a minute; weights on Hugging Face | |
| zembed-1 embeddings | 0.05 per million tokens | pricing page: weights available on request | |
| Search API: text extraction (OCR) | 1.75 per 1,000 pages | pay as you go | |
| Search API: indexing | 0.50 per MB | pay as you go | |
| Search API: storage | 0.10 per MB a month | pay as you go | |
| Search API: queries | 1.50 per TB searched | pay as you go | |
| Enterprise and on-premises | contract | volume discounts, deployment in the customer's cloud, 99.99% SLA |
All prices were usage-based: per million tokens of text, per 1,000 pages of text extraction, per megabyte indexed or stored, and per terabyte searched. Customers could prepay for higher rate limits. The page lists no monthly plans, so the yearly column is empty. Enterprise and on-premises deployment were priced by contract. New signups have been closed since 24 July 2026.
When did ZeroEntropy launch, and how did it grow?
From a 2024 start to a sale to its own customer, 14 months after the earliest dated description of the live product. Each entry carries its grade.
- 2024CGhita Houir Alami and Nicholas Pipitone found ZeroEntropy, with a pre-seed investment from Entrepreneur First. [4, 5]
- 14 May 2025CThe company describes "ZeroEntropy v1", a managed hybrid search engine that combines keyword search, embeddings and language models. [8]
- 9 and 10 July 2025CA US$4.2 million seed round led by Initialized Capital, the company says, and the release of the zerank-1 reranker at US$0.025 per million tokens through the API. [5, 6]
- 18 November 2025 to 2 March 2026CRelease of the zerank-2 reranker (18 November 2025) and of the zembed-1 embedding model (2 March 2026). [10]
- 24 July 2026ANotion announces that it has acquired ZeroEntropy and that the whole team joins. New signups close, and all models move to the Apache 2.0 license. [1, 2]
- 4 September 2026CThe last day of full support that the company announced; after it the products were to be shut down. [2]
Why did ZeroEntropy work?
Four causes, in the order they mattered. Each one is a reusable lesson.
Small specialized models
The company's thesis was that narrow, repeated steps in an AI product, such as ranking documents, do not need a large general model. Its about page claims that a small reranker tuned on a customer's own tests beats a frontier model at about 1% of the cost and 10% of the latency. Notion's own figures, 85% lower latency on the reranking step, are the strongest outside support for that claim.
Public benchmarks against larger providers
Releases came with comparisons on public and private test sets against Cohere, Jina, Voyage and a large language model used as a reranker, and the team published its training method (zELO) as a paper. For a developer choosing a vendor, these numbers were the sales material.
The customer became the buyer
Notion first used ZeroEntropy as a customer and replaced, in its own words, "offerings from larger and better-funded competitors". After about two months in production, it bought the company to hire the team that had trained the model.
How it was built
These points are the company's own statements on its blog and pricing page, grade C. [9, 8, 6, 11, 7, 2]
- zerank-1 and zerank-1-small were made by fine-tuning open Qwen models with 4 billion and 1.7 billion parameters, using LoRA, a method that trains a small set of extra weights instead of the whole model.
- Training labels came from language models that compared two documents at a time; an Elo rating, as in chess, turned the wins into one relevance score per document. Sampling about 400 pairs per query instead of 10,000 kept the cost down.
- The search engine of May 2025 turned every page into an image, read its text with OCR, described diagrams with a vision-language model, and stored keywords in ParadeDB and embeddings in Turbopuffer, and combined both kinds of search with language models.
- Models were sold through the company's API, with weights on Hugging Face and zerank-1-small also offered through Baseten; companies could deploy in their own cloud through AWS Marketplace and Azure.
- The pricing page claimed SOC 2 Type II, HIPAA, GDPR and CCPA compliance. Support ran through a Slack community, Discord and email.
Tools named in the sources
Each one opens its page in the Tools tab, with every story that uses it.
Who tried this before ZeroEntropy, and why did they fail?
Three companies in our database of failed startups sold search or company knowledge to businesses and failed. The values below are from that database and are not independently verified.
Omnity
The promiseSemantic search that connected data from many unstructured sources for research-heavy industries such as pharmaceuticals and law.
Why it diedLarge technology companies offered stronger AI search, and the cost of collecting and processing data was higher than the revenue.
Neeva
The promiseAn ad-free search engine paid for by a monthly subscription.
Why it diedToo few consumers paid for search. Snowflake hired the team and took the technology for enterprise AI search.
Niles
The promiseA Slack bot that answered staff questions from a company's own knowledge.
Why it diedSlack's own features and cheaper competitors left it no room.
The failed startups come from our database of 1,749 dead startups, retold here in our own words. Its entries are AI-assisted summaries of public sources, so these cards carry no evidence grade.
What changed
Omnity and Niles tried to sell a whole search product to end users and competed with large companies that offered search as one feature among many. ZeroEntropy sold one component, the ranking step, to developers who already had a search system, and charged by usage, so a customer could test it in an afternoon without a sales contract.
The end looks like Neeva's: a buyer took the team and the technology, and the product stopped. The difference is the buyer. Neeva's team went to a company that wanted its search skills; ZeroEntropy was bought by a customer that had already measured the model in its own product.
Can one person copy ZeroEntropy now?
The model work, hardly. The route to a buyer, yes.
Training rerankers and embedding models that beat larger providers needed a team of mathematicians and competitive programmers and about US$4.4 million of outside money, as the company stated. One person cannot easily copy that. The route is open to anyone with a narrow technical product: sell one measurable step to companies that build AI products, and let a large customer measure it in production.
Where it is still open: Other narrow steps that the company itself named: classification, routing, query rewriting and context compression for AI pipelines.
- 1Pick one narrow step in AI products that large general models do slowly or expensively.
- 2Publish benchmarks against the best-known providers, with the test sets.
- 3Sell through an API with usage prices and no sales call, so developers can try it the same day.
- 4Ask large customers for permission to publish their latency and accuracy results.
What could go wrong for ZeroEntropy?
The risks are the mirror image of the causes above.
What is not verified
No sale price, revenue figure or count of paying customers was found. We read Notion's announcement, ZeroEntropy's announcement and migration guide, its pricing, blog and about pages, and its Y Combinator page; none gives an amount. The team size of 6 comes from Y Combinator's directory, with no date. A reader could look for a post by the lead investor, Initialized Capital, or a later interview with the founders.
The product was shut down after the sale
New signups closed on 24 July 2026, and the company said all products would be shut down after 4 September 2026. Customers had to move to Cohere, Voyage or OpenAI, or host the now open weights themselves on Baseten or Modal, as the migration guide describes. The website was still online on 7 November 2026 with the notice; we did not test whether the API still answers.
Competition
Cohere, Voyage and Jina sold rerankers and embedding models to the same developers, and ZeroEntropy's own migration guide names Cohere, Voyage and OpenAI as replacements. On this site, Moss (a semantic search runtime for voice agents) and PDF Vector (a document parsing API) sell other parts of the same retrieval chain.
Dependence on model providers
ZeroEntropy trained its own models, so it did not pay a model provider per request. Its first rerankers were fine-tuned from open Qwen models, its training labels came from language models that the zELO post does not name, and its May 2025 search engine used language models and a vision-language model whose providers are not named.
Sources
The grade next to each source is the best grade that the source can support. "Read in full" means that the text of the page was downloaded and the numbers were checked in it.
- A Notion blog, 24 July 2026: ZeroEntropy is joining NotionRead in full. The buyer's own announcement. Source for the acquisition, the whole team joining, the founders' names, the new Model Research team and Notion's latency results.
- C ZeroEntropy, 24 July 2026: ZeroEntropy Joins NotionRead in full. The founder's post. Source for the closed signups, the shutdown date, the Apache 2.0 license and Notion as a top customer.
- C ZeroEntropy, 24 July 2026: ZeroEntropy Migration GuideRead the summary and the sections on models and deployment. Source for the replacement vendors and the self-hosting recipes.
- C Y Combinator company page: ZeroEntropyRead in full on 7 November 2026. Source for the founding year, batch, location, team size, founders' roles and the status "Acquired".
- C ZeroEntropy, 9 July 2025: ZeroEntropy Raises $4.2M Seed Round to Make AI Retrieval Truly IntelligentRead in full. Source for the seed round, the total raised and the Entrepreneur First pre-seed.
- C ZeroEntropy, 10 July 2025: Announcing ZeroEntropy's First Rerankers: zerank-1 and zerank-1-smallRead in full. Source for the first price, the benchmark comparisons, the licenses and Baseten.
- C ZeroEntropy pricing page, read 7 November 2026Read in full. Source for the price table and the compliance claims.
- C ZeroEntropy, 14 May 2025: Deep Dive: The Architecture of ZeroEntropy v1Read in full. Source for the search engine design: OCR, vision-language model, ParadeDB and Turbopuffer.
- C ZeroEntropy, 19 September 2025: Paper TLDR: How we trained zerank-1 with the zELO methodRead in full. Source for the pairwise training, the Elo scores and the Qwen base models.
- C ZeroEntropy blog index, read 7 November 2026Read the list of posts with their dates. Source for the release dates of zerank-2 and zembed-1 and the customer case studies.
- C ZeroEntropy home page, read 7 November 2026Read in full. Source for the acquisition notice, the named customers and the AWS Marketplace and Azure deployment.
- C ZeroEntropy about page, read 7 November 2026Read in full. Source for the company's claim about small models against frontier models.
- C Database of failed startups, local exportLocal export. AI-written summaries, not independently verified. Source for the failed predecessors.
Run this company? Send corrections or a verified figure to aiolum@livyden.com