Duolingo's 10-K defines the numbers blogs get wrong

Somebody told you Duolingo has 500 million users. Maybe you read it in a roundup, or a comment, or a slide.

Search Duolingo's own filings for it and you get nothing. The figures they do file are 133.1 million monthly active users, 52.7 million daily active users, and a definition of each that fits in one sentence.

I pulled four years of annual reports and ran the arithmetic, and the interesting part turned out to be the opposite of what I expected. Duolingo's own disclosure is careful to the point of being fussy. The mangling happens downstream, in pages that cite the company as their source. Every figure below comes from a document signed under penalty of perjury, with the definitions quoted verbatim so you can check my wording against theirs.

Two sentences that define everything

From the FY2025 Form 10-K, filed 27 February 2026 for the year ended 31 December 2025:

And the sibling:

Read what those cover. A monthly active user opened the app once in a calendar month, or loaded the learning part of the website once, and that counts. The filing does not claim they finished a lesson or say how long they stayed. Both definitions measure engagement with software, and the filing says so directly: MAUs "are a measure of the size of our global active user community on Duolingo," while DAUs "are a measure of the consistent engagement of our global user community."

Both figures are quarterly averages, which matters more than it sounds. The 133.1 million is an average of October, November and December 2025, and the 52.7 million is an average of 92 individual days.

Then the filing does something I did not expect. It tells you the numbers might be wrong, and how:

The same section says mobile background activity "can cause our system to miscount the user metrics." Every public company writes a version of that paragraph, so I will not oversell it as candour. But it is specific about the failure mode, and it means the strongest thing anyone can say about 133.1 million is that it is Duolingo's own unaudited estimate, disclosed as such.

Where the number loses a word

Here is where the 500 million comes from. Duolingo's 2021 registration statement, the S-1/A filed before the IPO, says: "With over 500 million downloads, our flagship app has organically become the world's most popular way to learn languages." A few pages later: "Our millions of learners complete over 500 million exercises every day."

Downloads, and daily exercises. Two different 500 millions in one document, neither of them a count of people.

I searched the S-1/A, all four annual reports from FY2022 through FY2025, and the most recent quarterly report for the phrase "registered users." It appears zero times, so any page quoting one is relabelling something else.

Take one live example, checked on 3 September 2026. A statistics page at gitnux.org, published 13 February 2026 and updated 20 June, listing investors.duolingo.com among its sources and carrying a "fact-checked" badge, states: "Duolingo has over 500 million total registered users worldwide." The same page states "Over 3 million users have achieved a streak of 365 days or more." The FY2025 filing says about 15 million.

Notice which way each error runs. The user count gets inflated roughly fourfold against MAU. The engagement figure gets deflated fivefold. Somebody spinning would spin both the same direction, so these are uncorrected rather than motivated.

Duolingo, for its part, leads with the smaller number. The Q4 2025 shareholder letter opens on passing 50 million daily active users, and the company's stated medium-term goal is "to reach 100 million DAUs in 2028." Both milestones are set in DAUs, which is the more conservative of the two metrics available to them.

Four years of filings in one table

I pulled every 10-K back to FY2022, which carries FY2021 as its comparative, and put the operating metrics in one place. All of these are quarter-four averages, taken from the filings themselves.

Duolingo Q4 operating metrics, from four annual reports

Q4 2021 42.4M / 10.1M MAU / DAU. Ratio 23.8%. 2.5M paid.

Q4 2023 88.4M / 26.9M MAU / DAU. Ratio 30.4%. 6.6M paid.

Q4 2025 133.1M / 52.7M MAU / DAU. Ratio 39.6%. 12.2M paid.

The gap between the two columns is the whole story of this post. In Q4 2025 it is 80.4 million.

Be careful with what that subtraction is, because I nearly got it wrong myself. MAU averages three monthly unique counts and DAU averages 92 daily ones, so they are differently built objects and 80.4 million is not a headcount of anybody. Somebody who used the app on one day of the quarter adds about a third of a person to the MAU average and roughly a hundredth to the DAU average. Treat the gap as a gap between two averages, which is still the most informative subtraction in the document.

Now the part that surprised me, and it runs in Duolingo's favour.

The ratio is climbing, hard. 23.8 percent in Q4 2021, then 26.9, then 30.4, then 34.7, then 39.6. Five consecutive years of the daily share of the monthly base going up. MAU grew 3.1 times across that span while DAU grew 5.2 times, so Duolingo is holding a larger fraction of its people day to day as it gets bigger.

Paid penetration moved the same way. Dividing paid subscribers by Q4 MAU gives 5.9 percent in 2021 and 9.2 percent in 2025.

And here is a trap I walked into, worth more to you than the number itself. Duolingo publishes its own penetration figure, built differently: "subscribers made up 9.2% of our average MAUs over the last twelve months as compared to 8.8% ... during the year ended December 31, 2024." Theirs is a trailing-twelve-month average, mine is a Q4 snapshot over a Q4 average. The two land on 9.2 percent for 2025 by coincidence, and disagree for 2024, where the filing says 8.8 and my division says 8.1.

Use the filing's series when you want penetration: 8.3, 8.8 and 9.2 percent across three years, theirs and consistently built. Mine only puts paid subscribers on the same Q4 basis as the rest of the table.

The quarterly reports carry it forward. The Form 10-Q filed 6 August 2026 puts Q2 2026 at 140.6 million MAU and 58.7 million DAU, up 10 percent and 23 percent on the same quarter last year. DAU is still growing more than twice as fast as MAU.

The streak figures, and what they count

Duolingo puts two streak figures in the business section of every 10-K, and they are the most useful numbers in the document, because a streak is the closest thing to a behavioural measure the company discloses. FY2025, verbatim: "As of December 31, 2025, there were about 43 million daily active users with a 7-day streak or longer, and about 15 million daily active users with a 365-day streak or longer."

Note the base. Both figures count daily active users, so they sit inside the 52.7 million rather than inside the 133.1 million. That distinction is the one the restatements lose.

Here is the arithmetic, which is the reason I wrote this page.

Q4 2025, my division of the filing's own figures

0

of daily active users hold a 7-day streak or longer

0

of monthly active users hold a 365-day streak

0

gap between the monthly and daily averages

About four in five daily users have a week-long streak going, which is better than I expected before I did the division. One caveat on that quotient: the 43 million is a snapshot dated 31 December, and the 52.7 million is a 92-day average, so 81.6 percent is an approximation rather than a clean ratio of like to like.

The year-long streaks have grown faster than anything else in the filings. About 2.5 million at the end of 2022, then 5 million, 10 million, 15 million. Six times in three years, against a 2.2 times rise in MAU, so the deep-habit population is growing about 2.7 times faster than the base it sits in.

As a share of monthly actives, the 365-day group went from 4.1 percent to 11.3 percent. Whatever you think of gems and leagues, the streak mechanic is doing what it was built to do, and the filed evidence for that is stronger than the filed evidence for anything Duolingo claims about learning outcomes.

Which is the honest boundary of this exercise. A 365-day streak means 365 days of completing at least one lesson, and the filing says exactly that, no more: "A streak indicates the number of days in a row a lesson is completed." It is not a proficiency measure, and Duolingo does not present it as one. For the argument about what daily app practice does and does not build, I made it separately in the piece on streaks and speaking.

What the ratio is worth to you

Three things I would take from a couple of hours in the filings.

First, when you see a Duolingo user number, ask which one. If a page says "users" with no qualifier, it has lost the distinction the source document spends two sentences establishing, and you cannot tell whether you have a monthly average, a daily average, a download count from 2021, or a number somebody invented.

Second, the comparison worth making for any app is DAU against MAU. Duolingo's 39.6 percent is high for a consumer app, and they publish it because it flatters them. Most apps never publish it at all.

Third, the one that touches you rather than the market. Whatever headcount sits behind that 80.4 million gap, it dwarfs the daily figure, so a monthly-but-not-daily relationship with a language app is the commonest relationship anyone has with one. The filings say nothing about what those people are doing, so I will not guess. It is still worth knowing that the version of yourself who studies twice a month is the statistical norm.

The engagement numbers say the streak works. The definitions say a streak counts days a lesson was finished. Both are true from the same document, and holding both at once is the whole skill of reading a filing.

I build Verbamor, so here is where it sits. It does nothing about the 80.4 million and does not compete for them. It does the narrow job of taking vocabulary out of lessons you already paid a human for, turning it into fill-in-the-blank cards with native audio, and scheduling them with FSRS, which is the loop in what happens to a word after a lesson. It is $19.99 a month, iOS only, and teaches you nothing new. With no tutor and no lessons there is nothing for it to hold, and a free Duolingo streak is a better use of your evening than a subscription that sits empty.

And if you will genuinely maintain your own deck, Anki is free and runs the same algorithm family once you enable it. I wrote the comparison against Anki without pretending the answer always comes out our way.

---

Sources

Every load-bearing claim Verbamor makes is traced to its paper on the research page.