Luis von Ahn grew up in Guatemala City going to a private English-language school, and he knew exactly why his mother had sent him there.
English was the ticket. Not just in Guatemala but across most of the world, being able to speak and read English changed what jobs you could get, what universities would accept you, what economic life was available to you. Von Ahn’s mother understood this well enough to pay for an expensive private school. Most Guatemalan families couldn’t. The people who needed English the most were the ones with the least access to it, and the main reason was money. A decent language course cost hundreds of dollars. A private tutor cost more. The software products like Rosetta Stone cost upward of $500 for a full program.
Von Ahn filed this observation away and went on to have a genuinely unusual academic career. He studied math at Duke, got his PhD in computer science at Carnegie Mellon, and became a professor there in 2006. Before Duolingo he had already built two products that Google bought. The first was CAPTCHA, the distorted-text test that websites use to tell humans from bots, which he co-developed with Manuel Blum in 2000. The second was reCAPTCHA, a cleverer version of the same idea: instead of showing users random distorted text, it showed them scanned words from old books and newspapers that computers couldn’t read accurately, and used the aggregate human responses to digitize those texts. Every time someone solved a reCAPTCHA, they were helping digitize the New York Times archive or a library of old books. Google bought reCAPTCHA in 2009. Von Ahn has said he sold it for “a lot of money.”
After the sale he was still teaching at Carnegie Mellon and started thinking about what to build next. The Guatemala observation came back. Language learning was expensive and the people who most needed it were the people least able to pay for it. He thought a free language learning platform could do something meaningful there. His graduate student at the time, Severin Hacker, who was Swiss and had grown up speaking Swiss German, thought so too. The two of them started working on it in 2009 and officially founded Duolingo in 2011.
The First Business Model, and Why It Had to Change
The original version of Duolingo was built around a clever economic mechanism that von Ahn had thought up. It was the same basic insight he had used with reCAPTCHA: make something useful to people and use the byproduct of their activity to generate value for someone else.
The idea was that people learning a language would practice by translating real content from the web. A company that needed documents translated, CNN or BuzzFeed or a law firm, would pay Duolingo for those translations. Duolingo would use that revenue to fund free lessons. Learners got free education. Paying clients got translations. The work the learners did to practice was the same work the clients needed done.
It was a genuinely elegant concept, and it didn’t work.
The translation model had two problems. First, translation quality from beginners is inconsistent. You can aggregate lots of translations and use statistical methods to get to something reliable, but scaling that process turned out to be operationally messy. Second, and more importantly, the learner experience that made for good translation practice did not make for good language learning. Teaching someone Spanish so they can accurately translate a CNN article about financial regulation is a different goal than teaching them to have a conversation or understand a movie. The two optimizations pulled in different directions.
By 2014 or so, Duolingo had largely moved away from the translation model and toward a straightforward freemium approach: the app was free, with ads showing up between lessons. If you wanted an ad-free experience and a few extra features, you paid for Duolingo Plus, which launched as a paid tier in 2017. The core lessons stayed free.
Von Ahn has been consistent about this ever since. The product that teaches people languages will always be free. The subscription tier is for people who want a better experience, not for people who want access to the thing at all.
300,000 People on the Waiting List Before Anyone Could Use It
When Duolingo launched its private beta in November 2011, over 300,000 people were already waiting for access. By the time the app opened to the public in June 2012, that number had grown to around 500,000.
This was not the result of a paid marketing campaign. Von Ahn had given a TED talk earlier that year explaining what Duolingo was going to be and why he was building it. The talk spread, people signed up for the waitlist, and by the time the app launched, there was already a substantial audience waiting.
The early version of the app had five languages: English, Spanish, French, German, and Portuguese. It was gamified from day one. Lessons were broken into short segments. You earned points for completing them. If you got too many wrong answers in a row, you lost a “heart” and had to restart the lesson. A daily streak tracked how many consecutive days you had done at least one lesson.
The streak turned out to be one of the most important product decisions in Duolingo’s history, though it didn’t look that way at first. Von Ahn has explained the origin of it directly: he and Hacker were trying to learn each other’s languages during the building of the app. Von Ahn was working on Spanish lessons for Hacker, a native Swiss German speaker; Hacker was building German lessons for von Ahn. Every morning von Ahn would come into the office and ask Hacker if he had done his Spanish lesson the day before. The answer was usually no. They needed some automated mechanism to create daily accountability. The streak was that mechanism.
By the end of 2024, over 10 million Duolingo users had maintained streaks longer than a year. One in five daily active users had a streak exceeding 365 days. A feature invented because two founders kept forgetting to do their homework became the psychological backbone of the product.
Building the Course Library Without Building All the Courses
One thing Duolingo got right early was recognizing it could not build every language course itself.
After the initial launch with five languages, users started asking for more. Obscure languages, minority languages, endangered languages, and in a few cases fictional ones. There was demand for courses in Klingon, in Navajo, in Hawaiian, in High Valyrian from Game of Thrones. Duolingo did not have the internal linguists or resources to build all of these.
In 2014, the company launched the Duolingo Incubator, a program that let volunteer contributors, native speakers, linguists, and enthusiasts apply to create courses using a tool Duolingo provided. The company would then review and launch the courses built by the community. This expanded the language catalog far beyond what a small internal team could produce, and it brought in communities of native speakers who had a stake in seeing their languages taught well.
The Incubator also created an interesting dynamic where some languages, like Irish and Welsh, had passionate volunteer contributor bases that produced unusually high-quality courses because the people building them cared deeply about the language’s survival. Duolingo became one of the largest platforms for learning Irish and Welsh, partly because Irish and Welsh speakers were motivated to build excellent courses.
At the same time, Duolingo was running massive amounts of A/B testing to improve the lessons that already existed. The company treated its user data the way a tech company should: as a signal about what was actually working versus what they assumed was working. They would test different ways of presenting new vocabulary, different frequencies for reviewing older material, different formats for exercises. The product improved continuously based on what the data showed rather than what linguists assumed was the right pedagogical approach.
This data infrastructure became a real advantage. Duolingo had more data on how people learn languages in an app than any other organization in the world, and they were using it.
The Green Owl and Its Reputation for Harassment
Duo, the green owl that is Duolingo’s mascot, has a complicated public image.
The character sends push notifications reminding users to do their daily lesson. Initially these were polite reminders. Over time, whether by design or by the emergent behavior of Duolingo’s marketing team, the notifications became more aggressive. Memes spread of Duo looking menacing. Jokes circulated that the owl would show up at your house if you missed a lesson. Duolingo’s marketing team noticed this cultural moment and leaned into it, putting out content that played up the owl’s passive-aggressiveness.
The result was something unusual: an educational app that became a genuine cultural presence. Users made fan art of Duo as a threat. TikTok videos racked up millions of views showing Duo appearing ominously in everyday situations. Duolingo’s social media team adopted an “unhinged” voice that matched the meme, posting content that was deliberately weird and self-aware.
Von Ahn has described this as Duolingo doing for marketing what it did for product: running experiments, noticing what worked, and doubling down. The Duo character’s viral reputation was not fully planned. It was noticed and then amplified. By 2024, Duolingo’s YouTube Shorts and Instagram Reels views grew over 430% and 450% year over yearrespectively, and AdAge named Duolingo its Marketer of the Year. The owl had become genuinely famous.
The English Test, and Selling to People Who Actually Had to Learn
One underappreciated part of Duolingo’s business is the Duolingo English Test.
The standard English proficiency exam for university admissions is the TOEFL. It costs around $250, takes about three hours, and has to be taken at a certified testing center, which can require travel. For someone in Guatemala, or in a rural part of India, or anywhere that doesn’t have convenient access to a testing center, this is a significant logistical and financial burden. Von Ahn experienced this himself: he had to fly from Guatemala to El Salvador to take his English proficiency exam when applying to American universities.
Duolingo’s version of the test costs $49, takes 45 minutes, and can be taken from home. By 2021, over 3,000 institutions worldwide, including Duke, UCLA, Columbia, Dartmouth, NYU, and Yale, accepted the Duolingo test in their admissions process. The test generates around 10% of Duolingo’s total revenue and addresses a genuinely different population than the app’s casual learners: people who need to prove English proficiency to get into a university, get a visa, or qualify for a job.
The Duolingo English Test is not a gamified experience. There are no streaks or points. It is a real standardized test that competes directly with TOEFL and IELTS on price, convenience, and increasingly on acceptance. It represents the other side of what von Ahn built Duolingo to do: not just make language learning accessible, but make the credentialing around language learning accessible too.
Going Public and Where the Numbers Landed
Duolingo went public on the Nasdaq on June 28, 2021, under the ticker DUOL, at a valuation of $3.7 billion. By the end of the first trading day, the market cap was sitting around $5 billion.
At the time of the IPO, the app had around 500 million registered users and roughly 40 million monthly active users. Revenue was $161 million in 2020, growing fast. The subscriber penetration rate, the percentage of monthly active users paying for a subscription, was just above 5%.
What happened over the next three years was that all of those numbers moved in the right direction at the same time, which doesn’t always happen. Daily active users went from around 8 million at the 2021 IPO to 40.5 million by the end of 2024, a 51% increase in a single year. Monthly active users hit 116.7 million. Paid subscribers reached 9.5 million, up 43% year over year. Subscriber penetration climbed from 5% to around 8.5%.
Revenue reached approximately $748 million in 2024. Operating margins went from deeply negative to positive: the company posted operating margins of around 8.4% in 2024 after years of losses. The free cash flow margin in Q4 2024 was 42%.
Duolingo launched Duolingo Max in 2023, a higher-priced tier built around GPT-4 that offered features like Explain My Answer, where users could ask why their answer was wrong and get a detailed explanation, and Roleplay, where users practice conversation with AI characters. Max is priced at around $168 per year versus $80 per year for the basic Super Duolingo tier. By late 2024, Max was in 5 courses across 27 countries and was described by von Ahn as exceeding expectations.
The AI push extended beyond subscription features. In 2024, Duolingo published 7,500 content units using AI-assisted creation tools, up from 425 in 2021. The company was using AI to write exercises, create audio, and generate practice content at a scale that would have been impossible with only human content creators.
What von Ahn Still Filters Every Decision Through
In interviews over the years, von Ahn has returned to the same framing when asked how Duolingo decides what to build.
He asks whether the feature would help someone in Guatemala learn English.
It is not a complicated filter, but it has held for fifteen years. When the company debates adding a feature that only works well on expensive hardware, that filter pushes back. When someone suggests a paywall for core learning features, the filter rejects it. When the product team looks at which languages to prioritize, the filter points toward the ones that create the most economic opportunity for learners in developing countries, which is why English-learning courses get disproportionate resources even though Duolingo users learning English are not the company’s most profitable customers.
By early 2025, Duolingo had 40.5 million daily active users, 116.7 million monthly active users, and 9.5 million paid subscribers. Nearly 2 billion people worldwide are estimated to be learning a language at any given time. Von Ahn’s stated position is that Duolingo has barely scratched the surface of that market.
The company started because a Guatemalan kid noticed that English courses were too expensive for most people in his country. The kid grew up, invented CAPTCHA, sold reCAPTCHA to Google, became a MacArthur Fellow, and then spent the next fifteen years building the thing he thought the world actually needed. As of 2025 it has been downloaded more than any other education app in history and generates nearly $750 million a year while keeping its core product free.
Leave a Reply