Duolingo: becoming a vertical AI
$DUOL Q4 2025 ER Update
The content of this analysis is for entertainment and informational purposes only and should not be considered financial or investment advice.
Duolingo just closed a strong FY 2025: revenue grew 39% YoY to $1B+, but user growth is clearly slowing, which will likely slow revenue growth next. QoQ MAU is negative, and the average over the last three quarters is in the low single digits, which suggests DAU will follow, and revenue right after. This is exactly what management guided: 10–12% bookings growth for 2026.
This was mostly expected, at least to me. The more interesting news is that Luis von Ahn, CEO and co-founder, announced a strategic fork in the road, aimed to reach 100M DAU in 2028.
He basically said:
AI is reshaping learning, so in 2026 we will deliberately prioritize user growth and teaching better, even if that moderates near-term financial growth.
But I want to go one step further.
My thesis: the next obvious move for Duolingo is to train a great open LLM using their unique data, and turn it into the best teacher in the world.
Not a chatbot. A teacher.
And I think the timing is perfect: the next 12 months are when open models will likely become “smart enough” to deliver consistent teaching quality at low cost. Let’s get into it.
Table of Contents
Q4 2025 Update
L2C take aways and performance
Business Ontology Framework by L2C
Business Ontology
4D Valuation Model
L2C portfolio strategy
Q4 2025 Update
Duolingo is quietly building the most valuable dataset in education
The reason Duolingo is special is not “language learning”.
It’s this:
They have a massive funnel (free product).
They have daily repetition behavior (habit).
They have outcomes that can be measured (did you answer correctly? how fast? how many retries? do you remember it next week?).
That is the dream dataset for training a teaching model.
Think of it like self-driving:
The car company that collects the most real-world driving data wins.
Education is similar:
The platform that collects the most real-world learning data wins.
The real moat is not the content library.
It’s the feedback loop.
And Duolingo’s feedback loop is insane.
They already use AI… but the current approach is limited
Duolingo has used ML for years (e.g. personalization models like Birdbrain, and AI-assisted content generation).
More recently, they productized GenAI in Duolingo Max: Roleplay and Video Call with Lily, powered by modern LLMs.
And if you read how Video Call works, you realize they’re already doing “teacher engineering”:
They use the transcript to extract a “list of facts” about the user
Then they feed it back as system context to make future sessions more personal
They carefully prompt the model to ask the right question at the right level
This is not a toy.
It’s the early version of a learning agent.
But today there’s a problem:
The best teaching experience is gated behind expensive closed models (or expensive inference).
That makes it hard to scale “teacher quality” to hundreds of millions of people.
So you get a dilemma:
Either keep AI features premium-only (good for ARPU, bad for data + adoption)
Or democratize them (good for growth, expensive, margin pressure)
Management is basically choosing the second path in 2026: grow the user base and improve teaching quality, even if it hurts near-term numbers.
That choice makes perfect sense if they’re playing a longer game.
The longer game: train an open teacher model
Here’s the hot take.
Duolingo should fine tune and optimize an open model for one thing:
Teaching.
Not general chat.
Teaching.
Why open?
Because the economics of education are brutal.
If you want to serve 100M+ daily users with interactive speaking, explanations, coaching, and personalization, you need:
Low inference cost
High controllability
Strong safety / guardrails
Ability to run custom fine-tunes + retrieval
A model you can iterate on every week
Owning your own model is the only way to make that sustainable.
And “open” is the fastest path to reach that point, because:
Open models are improving at an insane pace
The ecosystem is global (research + tooling + fine-tuning stacks)
You can start from a very strong base model and then specialize
The missing ingredient is high-quality domain data.
Duolingo has it.
What data do they have that others don’t?
Most LLMs learn language from the internet.
Duolingo has something different:
Paired learning data: a prompt, an answer, and whether it was correct
Difficulty labels (implicitly via user performance)
“confusion patterns” (what people get wrong, systematically)
longitudinal memory: did you retain it after 7 days / 30 days?
speaking practice data (especially as Video Call expands)
This is a goldmine.
If you train a model with that feedback loop, you don’t just get a model that “speaks Spanish”.
You get a model that knows how to teach Spanish.
That’s a completely different product.
The KPI nobody tracks: words spoken per user
Management actually called this out on the Q4 2025 earnings call. Luis said that if you look at the “graph of words spoken per user on Video Call”, it’s been “a monotonically increasing graph over time.”
That’s not a random metric.
And the analogy is straight from Elon.
Tesla didn’t win autonomy by optimizing for “time spent in the car”. They obsessed over miles driven, because miles driven = real-world training data. More miles means faster learning loops, edge cases, and better models.
Duolingo’s version of “miles driven” is words spoken.
Because the endgame is not “minutes spent”.
It’s “practice delivered”.
If Duolingo becomes the best teacher in the world, the biggest unlock is speaking.
Speaking is where:
learners freeze
tutors are expensive
scheduling is annoying
and progress is slow
An AI teacher solves this by making speaking practice:
instant
infinite
personalized
cheap
So I expect the Duolingo product roadmap to push more and more toward:
conversation
pronunciation
adaptive feedback
“live” tutoring
Video Call with Lily is the prototype.
The next step is to make that prototype scale to everyone.
Why I think this happens in the next 12 months
Two reasons:
Open models are getting good enough
The last 24 months were about “wow, this can talk.”
The next 12 months are about “wow, this can reliably follow instructions, be consistent, and do narrow jobs extremely well.”
That’s when a specialized teaching model becomes viable.
Duolingo is explicitly shifting to a growth + teaching year
They are telling you: we are choosing the long-term user base over near-term monetization.
This is exactly what you do when you’re about to unlock a new growth engine.
What does Duolingo become if this works?
If you believe this path, Duolingo stops being “a language app”.
It becomes:
the default interface between humans and learning
Like Google is the interface between humans and information.
Or like Spotify is the interface between humans and music.
The big idea is simple:
if you own the best teacher, you own the learning time
and learning time expands with AI (because the cost goes down)
This is why I think management’s 100M DAU ambition is not crazy — it’s a stepping stone.
L2C take aways and performance
Q4 numbers were weak — but the story is bigger than a quarter.
Duolingo is positioning itself for the next era of learning, where:
teaching quality is AI
practice is unlimited
cost is low
and the best teacher wins
The obvious next step is to stop renting intelligence… and start owning it.
Train the teacher model.
Make it the best in the world.
If they execute, Duolingo won’t just be an edtech company.
It will be an AI company whose product is human progress.
As always, here is the “Deep Dive To Date” (DDTD), that is how the DUOL stock is performing since my initial deep dive on the January 11th 2025, when the price was $318.15.
-70% DDTDFrom here on, the content is restricted to L2C Premium Members, folks who’ve chosen to unlock this toolkit and support my independent research:
Business Ontology Framework by L2C
Business Ontology: My core blueprint for modeling and tracking company performance at every level.
4D Valuation Model: the valuation tool I use to value all my investments (Fair price is useless)
L2C Portfolio Strategy: My portfolio allocation and strategy in details
L2C Portfolio access & trades alerts: Real-time views into my holdings, plus instant notifications on buys, sells, and shifts.
Business Ontology Framework by L2C
The Business Ontology is a framework I built after tearing apart several tech companies from the ground up—breaking them down to their basic parts and piecing together a real thesis on what drives them. Think of it as a map of a company’s soul. It’s a tight set of core indicators—tailored to each business—that show if it’s heading the right way, no matter what the stock price says. These aren’t your basic stats like P/E ratios or revenue bumps you grab from Yahoo Finance. They’re deeper, sharper, and linked straight to the thesis I’ve cooked up on how the company makes value and fights in its market.
The framework boils down to two big pieces:
The Business Ontology—This checks if the company’s worth buying into or hanging onto.
The 4D Valuation Model—This gives you a 4D roadmap to guide your own calls on how much to put in (allocation).
Now, let’s dive into Duolingo and this quarter’s update.
