Flua

An attempt at the most efficient way to learn a language

How I built Flua's decks, step by step, and the research behind each choice.

Flua is a language-learning app I built around one question: what is the least time an adult can spend to understand a new language, if every minute goes where it buys the most understanding? My answer is a beginner course for each of seven languages (French, Spanish, German, Italian, Japanese, Chinese and Korean), taught as one deck of sentence flashcards. You press Start, hear a short sentence, try to recall what it means, and tap Good if you knew it or Again if you didn’t. Everything else is in how the deck is built, held to rules every card must pass, and the reasoning behind those rules is short enough to follow from start to finish.

Learning a language is a sum you can shrink

The goal here is the first big one: understanding everyday spoken sentences, at the speed people talk, which is where speaking, reading and everything else build from. That means recognizing the words and patterns people actually use. So the time it takes comes down to four numbers: how many words and patterns you need before everyday speech makes sense, how long each one takes to learn, how long it takes to keep, and how much time goes to anything that is none of these. A method gets faster only by shrinking one of those numbers without growing another.

Each number has a research-backed way to shrink it; how much the combination saves is what Flua measures. None of the ideas below is new. What I couldn’t find anywhere was a deck built to hold all of them, on every card, at the same time, and that turns out to be far harder than any one of them.

Each part has a known way to shrink it

The first number is how many words and patterns you need. You need fewer if you learn them in the order people actually use them. Words are not equally useful. In corpus studies, the thousand most common word families make up roughly 80 to 90% of the words in everyday speech 1, and the same steep curve holds in every language. So every Flua deck is ordered by how often each form is actually heard, measured from film and TV subtitles: 313 million words for French and 413 million for Spanish. In French, Spanish and German, the forms taught in a beginner deck cover around three-quarters of all the words spoken on screen; every deck’s live figure is on its transparency page. That is the steep part of the curve. The tail is long: following speech with ease takes several thousand word families 1, so a beginner deck is built to buy the first three words in four as cheaply as it can, and the later levels are for the rest.

I count exact forms rather than word families because a learner has to learn each form they can’t predict. Hearing “voyez” doesn’t tell a beginner it comes from “voir” (to see). If every unpredictable form and every meaning is treated as its own concept, each card can carry exactly one; if they’re lumped together, a card quietly teaches two things at once. Forms you can predict cost nothing extra: once you know “avais” (I had) and the card on person endings, “avait” (he had) is yours without a card of its own. One new concept per card is the rule every deck is built around, and every rule after it makes that one harder to keep:

  • Every meaning separately. French “nous” means “we”, “us” and “to us”: three cards. Spanish “si” and “sí” get five, from “yes” and “if” to the annoyed “but” of “¡Si ya te lo dije!” (but I already told you!). No sentence uses a word in a meaning you haven’t met.
  • Phrases whose meaning isn’t the sum of their words. I counted every one- to five-word sequence in each language’s subtitles and pulled out the frequent ones that mean something their words don’t. “Ça va” is literally “that goes” and means “how’s it going”; “qu’est-ce que” is literally “what is it that” and simply means “what”. Across the seven languages, 599 phrases like these turned up; the ones a beginner deck needs are taught as units, and each deck’s count is on its transparency page.
  • Audited sources. The French ranking matches an independent film-subtitle reference at a rank correlation of 0.938. A widely used Japanese frequency list was thrown out after the audit found it had merged words that sound alike, turning しない (“don’t do”) into 竹刀 (“bamboo sword”).

The second number is how long each concept takes to learn. That time drops when the concept is the only unknown in front of you and its meaning is shown. Pairing a word with its translation is one of the most effective ways studied to learn what it means 2, and beginners struggle to attend to form and meaning at the same time 3. So each Flua card teaches exactly one new concept, a word, phrase or grammar pattern, inside a sentence whose every other word you already know by exact form and meaning.

Grammar follows the same rule. People often ask whether flashcards can teach grammar. Each language has a map of its grammar patterns, 120 for French. The most basic ones, like articles and the present tense, are in nearly every sentence from the first cards, so they are learned from use and get no card; in French that’s 37 of them. Every other pattern gets one card, placed before the first sentence that needs it, so the beginner deck teaches the ones its sentences use and later levels will bring the rest. Its note is short and uses only the card’s new word and words you already know, where a dictionary or textbook explanation would be full of words you don’t. Across the seven decks, about one card in 28 is a grammar card. Adults learn grammar better with explicit explanation than from exposure alone 12, and practice in understanding sentences builds grammar about as well as practice in producing it, at least in the short term 11. So Flua explains each pattern once and lets hundreds of sentences do the rest, with no tables and no drills. You can use the present perfect correctly without ever having learned its name.

Speaking follows the same rules. You can’t say a word you haven’t learned, so Flua teaches you to recognize each one first. Then, in the iPhone and Android apps, it turns its sentences around. A speaking card shows you the meaning, you say the whole sentence, then you check it and rate yourself. Recognizing a word and producing it each improve most with practice in that direction 16, so every speaking card is scheduled as its own memory, and it never arrives before you know every word in its sentence. Speaking cards follow the deck’s order, so the first things you learn to say are the things people say most. You set how much of each session goes to speaking, from none to about one new speaking card for every new sentence.

The third number is how long it takes to keep what you’ve learned. That time drops when each review comes at the right moment and takes only a few seconds. Spaced retrieval is one of the best-replicated findings in learning research 4, 5. Flua schedules every card with FSRS-6, the open memory model also built into Anki, which predicted forgetting more accurately than Anki’s older algorithm for about 99.5% of 10,000 users in an open benchmark 6. Reviews are built to be quick: sentences are at most ten words, about five on average, and there are two buttons, Again and Good. One quick decision per card keeps a rhythm going, and a session settles into flow.

Every review also trains the ear, since recognizing a word in speech is a partly separate skill from knowing it on paper 7: the full sentence plays at natural speed, voiced with the emotion of its scene, and each word comes back later in a new sentence and often a new voice 8, 9. Japanese, Chinese and Korean cards show a pronunciation line under the native script, so a beginner can start before they can read it.

The last number is time spent on anything else, and in Flua it is close to zero. There are no courses to choose, decks to find or cards to write, no settings to tune and no game screens between cards. Writing your own deck, which many serious learners do, is the biggest of these costs: at even a minute a card, a thousand cards add up to about 17 hours of making cards, on top of studying them, and holding the rules above by hand means checking every new sentence against every word and meaning already in the deck. The research and the work are in the deck, not on you. Flua keeps a streak, milestones and a daily reminder, because an intact streak brings people back 10 and coming back is what spaced review depends on. It never threatens you for missing a day, and there are no points, leagues or gems. Flua is built for people who have decided to learn a language and want their minutes to count.

The hard part is doing all of them at once

These rules fight each other. The most natural sentence for a common word usually contains a word the learner hasn’t met yet. One new concept per card narrows the sentences available; short and natural narrows them again. And every card limits what every later card may say. With about 1,500 candidate sentences for a typical French word, the possible ways to build a 1,000-card deck run to thousands of digits, and almost all of them break a rule somewhere. Solving for a single deck can take days of computation.

Here is what holding every rule looks like on one card, number 53 in the French deck:

Elle a dit quoi ? (What did she say?)
  • One new concept: “dit” (said). It is the only form on the card the learner hasn’t studied.
  • Every other word known: “elle” (she) from card 5, “quoi” (what) from card 15, and “a” (has) from card 52. English builds this past tense the same way (“she has said”), so the literal reading, “She has said what?”, already makes sense, and the card’s note names the pattern.
  • Frequent: all four forms are among the 75 most common in French film and TV.
  • How people actually talk: spoken French puts the question word last. The textbook “Qu’a-t-elle dit ?” would need forms the learner hasn’t met, and few people say it.
  • Short: four words, a three-second review.
  • Grammar just in time: a one-line note explains “a + dit = said”, the “has + done” pattern French uses to put many verbs in the past.
  • It comes back: “dit” returns in new sentences on cards 83, 95, 105 and 117, such as “Je vous ai dit ça, non ?” (I told you that, didn’t I?).
  • Voiced and pictured: the whole sentence at natural speed, with expression matched to its scene, and a picture of that scene.

That is one card. A deck is about a thousand of them, and every deck is held to the same 21 checks, ranked under six principles, with the one-new-concept rule above all others. Each deck’s measurements, from coverage and sentence length to grammar spacing, are published at tryflua.com/transparency, including the ones it doesn’t meet yet.

Every language is hard in its own way

The same rules meet different problems in each language:

  • French nouns have a gender you can’t guess. So each noun arrives with a word that shows its gender, usually its article, in a natural sentence (“la table”, “le pain”), and the gender rides along with the word.
  • German splits some verbs in two. “Aufmachen” (to open) shows up as “Mach die Tür auf”, with “auf” at the far end, so the deck teaches split verbs as one meaning across the gap: “Er sieht OK aus” (he looks OK) teaches “sieht ... aus” as “looks”.
  • Spanish has two verbs for “to be”. “Soy” and “estoy” both mean “I am”, and choosing between them follows rules a beginner can’t guess, so the deck teaches the contrast as its own card the moment a sentence needs both.
  • Japanese stacks endings onto verbs. One verb can take dozens of forms, so the deck teaches each common ending once as a pattern that works across the verbs it applies to, and teaches irregular forms on their own.
  • Korean glues particles onto words. “그가” is “he” plus a subject marker, so here the deck does the opposite of French: it splits the pieces, teaching the word and the marker once each, so you recognize them everywhere. The right unit is chosen per language.
  • Chinese has no spaces between words, and small words do several jobs. The deck checks that no compound appears before you’ve learned it, even when you know each character, and gives 了 two cards because it does two different jobs: “我说了” (I did tell them) and “我不想去了” (I don’t want to go anymore).

Whatever you need the language for

People learn a language to move abroad, to work, to pass an exam or to travel, and each of those adds words of its own: a lease, a field’s jargon, an exam’s topics. Underneath, they all start from the same core. The most common words and patterns make up most of what anyone hears and says in almost any setting, which is why vocabulary research advises learning them first, whatever the goal, and the specialized words afterward 1, 17. That core is the part Flua is built to get you through in the least time.

A move to France still takes conversations with people and the words for your own paperwork; an exam such as the DELF, TCF or TEF has its own format and writing; a job has terms you’ll mostly learn on the job. Each of those goes further once the everyday language is already yours, so whatever the goal, the first step is the same.

Watching shows teaches more once you know the common words

A common answer to “how do I learn a language fast?” is immersion: watch TV and films in it. Watching does teach words without any studying, and it is worth doing. How much it teaches depends on how many words you already know.

In studies where learners watched video without captions and without being told they would be tested on words, they picked up, on average, about 7% of the unfamiliar words tested, and still knew about 5% on a later test 18. With word-focused activities such as flashcards and word lists, learners recalled about 60% of the words right afterward and about 40% on a later test 2. The studies differ, so those figures don’t compare exactly, but the gap is wide. What you already know matters as well. University students at an intermediate level who watched an hour-long documentary learned about four new words from the 64 tested, about 8% of the ones they hadn’t known, and the more words a learner already knew, the more likely they were to learn new ones from the same hour 19. Over a school year of watching shows, gaming and reading online, the words teenagers already knew predicted how many new ones they picked up better than how much of those things they did 20. Most of what is said on TV is made of common words: the 3,000 most common word families, plus names, cover about 95% of the words in typical programs, and the researchers who measured that concluded that a learner who knows them and watches at least an hour of television a day has the potential to learn a significant amount of vocabulary from it 21.

So the order matters. The words shows repeat most are the common ones a beginner deck teaches on purpose, and learning them first makes every hour of watching count for more. Watching works best alongside the deck: study every day, and watch something you enjoy in the language, with captions in the language, which help both understanding and word learning 22, and ideally episodes of one series, which repeat more of the same words than unrelated shows do 23. You don’t need to watch all day. Early on the deck does most of the work, and as the words you know add up, each episode teaches you more.

A number that’s hard to flatter

How much of a language do you understand? Most learners can’t say. Streaks count days, not understanding, and the feeling of progress is a poor guide: in learning research, people regularly judge the easier way of studying as more effective even when the harder way leaves them knowing more 13, 14. Flua can feel like that at first. Recalling what a sentence means before you see it takes more effort than picking from options or matching tiles, so a session can feel harder and slower than a game, even while it works. That’s why Flua shows you numbers instead of asking you to trust the feeling.

So I built the Flua Index. It takes a large sample of everyday sentences and estimates, from the size of your vocabulary and the order words are most often learned in, the share you could follow word for word. “Learned” means more than seen: each card counts in proportion to how firmly the memory model predicts you’ll remember it. A sentence with even one word you don’t know doesn’t count, because one unknown word can be enough to lose what a sentence means.

It is a strict measure of vocabulary coverage rather than a comprehension test, and that is what makes the number honest. Seven common Japanese words make up about 11% of everything spoken, but knowing them doesn’t mean you understand 11% of Japanese; not one sample sentence is made entirely of them, and the Flua Index says 0%. It also makes progress concrete: a share of everyday sentences you can follow says more than “I’m B2” usually does.

Why I built Flua

About a month before a solo trip to Japan, I started Duolingo and took it seriously: half an hour every day for thirty days. It soon felt repetitive, so I started reading everything I could find about how people actually learn languages fast. That led me to Anki and its thousands of shared decks. Within twenty minutes on a popular one I could tell I was learning far faster per minute than I had all month, even though one of its first ten words was “tsumetai”, “cold to the touch”. (Flua won’t teach you that one early.)

Then I found a Japanese deck that changed how I thought about all of it. Every card was a line from a real anime with a screenshot of the scene, every sentence added exactly one new thing, and grammar lessons were woven in the same way, so I never studied grammar separately. It was built largely by hand and only loosely followed how common each word was, but that one rule made the difference. I barely used it in Japan, maybe half an hour a week, and people were still surprised by how much I could say. Back home I studied it properly for thirty days, half an hour a day, about 1,500 cards, and that is where most of my Japanese came from. I never took an exam and never learned to read or write, so I can’t give you a level; Flua’s placement test is how I’d measure it now.

A year later, planning a trip to France, I looked for the same thing in French and couldn’t find it: not even a decent deck that held to one new thing per card, and app after app with some flaw that made learning feel painfully slow. I’m a software engineer (I studied computer science at Carnegie Mellon, which, funnily enough, is also where Duolingo began), so I thought: how hard can it be? French has no equivalent of the huge, fan-subtitled anime catalog that Japanese deck was cut from, so the deck had to be built differently, then it needed an app that was pleasant to use, and every improvement revealed another. That has taken about two years. The deck that taught me Japanese held one rule; every Flua deck is held to that rule and twenty other checks.

These days I rarely find a way to make a deck more efficient. My belief, and it is only a belief until enough people use Flua to test it, is that each deck is close to the best balance these competing rules allow. If that’s right, the beginner stage of a language takes weeks of half-hour days rather than years, and the forecast at tryflua.com/transparency, which moves with real learners’ pace, will show how long real learners take to get through each level.

What to pair it with

  • Conversation still takes people. Speaking cards in the phone apps train you to say sentences from their meaning. Holding a real conversation, at another person’s speed and with their questions, still takes practice with people; Flua’s job is to make sure you have the words when you do.
  • Levels. Every language starts with its beginner deck, and I’m extending each one level by level toward C2 as quickly as I can; the current size of every deck is on its transparency page. A lifetime purchase includes each new level.
  • Your own cards. Flua’s decks can’t be edited and there is no Anki import. For flashcards on other subjects, or to keep reviewing cards you’ve already made, Anki is the better tool.

Test it on yourself

You don’t have to take any of this on trust. After every session, Flua shows your own measured pace (seconds of study per card, recall, words learned) and a forecast of how long the next level will take you.* An hour is enough to see your own pace.

Flua’s Progress screen: 13% of the way to A1, 126 of 1,000 words learned, 252 cards seen, 2 hours 12 minutes studied, and the forecast to A1Flua’s Progress screen: 13% of the way to A1, 126 of 1,000 words learned, 252 cards seen, 2 hours 12 minutes studied, and the forecast to A1
A real Progress screen: 126 words learned in 2 hours 12 minutes of study. Its forecast, 17 hours of study or 35 days at 30 minutes a day, blends Flua’s default pace with this learner’s; the measured pace alone works out to about 15 hours.

If you’ve studied a language before, the placement test in Settings starts you at your level and gives you your Flua Index on day one. It isn’t fooled by one forgotten word, and every word it marks as known is quietly re-checked in your reviews over the following weeks, so a wrong placement corrects itself. Every card in every language, more than 7,000 of them, is free right now, and you don’t need an account to start.

The same numbers, pooled across signed-in learners, are published live at tryflua.com/transparency. Controlled, independent outcome studies are rare even for the largest apps: Duolingo’s published outcome studies are company-run or company-funded, and its 2024 report tested self-selected course finishers with no comparison group 15. Flua’s answer is to show every learner their own numbers, and everyone the pooled ones.

Everything in Flua is free right now, and every card you study while it’s free and sync to your account before the free period ends stays unlocked for good. After that, the first 150 cards of every language stay free. Owning one language costs $89.99 once, including every future level, or $39.99 a year for that language; all languages together are $14.99 a month.

\* The forecast starts from a default pace and moves toward yours as you study. Without that anchor, someone who already knew a little of the language and raced through the first few dozen cards would be told they’d reach the next level in half an hour.

To try it, go to tryflua.com; the comparisons with Duolingo and Anki and common questions are there too.

Sources

  1. Nation, I. S. P. (2006). How large a vocabulary is needed for reading and listening? Canadian Modern Language Review, 63(1), 59-82. https://doi.org/10.3138/cmlr.63.1.59
  2. Webb, S., Yanagisawa, A., & Uchihara, T. (2020). How effective are intentional vocabulary-learning activities? A meta-analysis. Modern Language Journal, 104(4), 715-738. https://doi.org/10.1111/modl.12671
  3. VanPatten, B. (1990). Attending to form and content in the input. Studies in Second Language Acquisition, 12(3), 287-301. https://doi.org/10.1017/S0272263100009177
  4. Cepeda, N. J., et al. (2006). Distributed practice in verbal recall tasks. Psychological Bulletin, 132(3), 354-380. https://doi.org/10.1037/0033-2909.132.3.354
  5. Kim, S. K., & Webb, S. (2022). The effects of spaced practice on second language learning: A meta-analysis. Language Learning, 72(1), 269-319. https://doi.org/10.1111/lang.12479
  6. Open Spaced Repetition, srs-benchmark (about 10,000 Anki collections). https://github.com/open-spaced-repetition/srs-benchmark
  7. Milton, J. (2013). Measuring the contribution of vocabulary knowledge to proficiency in the four skills. EUROSLA Monographs 2, 57-78. https://www.eurosla.org/monographs/EM02/EM02tot.pdf
  8. Webb, S. (2007). The effects of repetition on vocabulary knowledge. Applied Linguistics, 28(1), 46-65. https://doi.org/10.1093/applin/aml048
  9. Barcroft, J., & Sommers, M. S. (2005). Effects of acoustic variability on second language vocabulary learning. Studies in Second Language Acquisition, 27(3), 387-414. https://doi.org/10.1017/S0272263105050175
  10. Silverman, J., & Barasch, A. (2022). On or off track: How (broken) streaks affect consumer decisions. Journal of Consumer Research, 49(6), 1095-1117. https://doi.org/10.1093/jcr/ucac029
  11. Shintani, N., Li, S., & Ellis, R. (2013). Comprehension-based versus production-based grammar instruction: A meta-analysis. Language Learning, 63(2), 296-329. https://doi.org/10.1111/lang.12001
  12. Norris, J. M., & Ortega, L. (2000). Effectiveness of L2 instruction: A research synthesis and quantitative meta-analysis. Language Learning, 50(3), 417-528. https://doi.org/10.1111/0023-8333.00136
  13. Kornell, N., & Bjork, R. A. (2008). Learning concepts and categories: Is spacing the "enemy of induction"? Psychological Science, 19(6), 585-592. https://doi.org/10.1111/j.1467-9280.2008.02127.x
  14. Deslauriers, L., McCarty, L. S., Miller, K., Callaghan, K., & Kestin, G. (2019). Measuring actual learning versus feeling of learning in response to being actively engaged in the classroom. PNAS, 116(39), 19251-19257. https://doi.org/10.1073/pnas.1821936116
  15. Jiang, X., Hopman, E., Reuveni, B., & Kittredge, A. (2024). Duolingo Path Meets Expectations for Proficiency Outcomes (Duolingo Research Report DRR-24-04). https://duolingo-papers.s3.amazonaws.com/reports/Duolingo_whitepaper_language_read_listen_write_speak_2024.pdf
  16. Webb, S. (2009). The effects of receptive and productive learning of word pairs on vocabulary knowledge. RELC Journal, 40(3), 360-376. https://doi.org/10.1177/0033688209343854
  17. Nation, I. S. P. (2013). Learning Vocabulary in Another Language (2nd ed.). Cambridge University Press.
  18. Webb, S., Uchihara, T., & Yanagisawa, A. (2023). How effective is second language incidental vocabulary learning? A meta-analysis. Language Teaching, 56(2), 161-180. https://doi.org/10.1017/S0261444822000507
  19. Peters, E., & Webb, S. (2018). Incidental vocabulary acquisition through viewing L2 television and factors that affect learning. Studies in Second Language Acquisition, 40(3), 551-577. https://doi.org/10.1017/S0272263117000407
  20. Laufer, B., & Vaisman, E. (2023). Out-of-classroom L2 vocabulary acquisition: The effects of digital activities and school vocabulary. Modern Language Journal, 107(4), 854-872. https://doi.org/10.1111/modl.12880
  21. Webb, S., & Rodgers, M. P. H. (2009). Vocabulary demands of television programs. Language Learning, 59(2), 335-366. https://doi.org/10.1111/j.1467-9922.2009.00509.x
  22. Montero Perez, M., Van Den Noortgate, W., & Desmet, P. (2013). Captioned video for L2 listening and vocabulary learning: A meta-analysis. System, 41(3), 720-739. https://doi.org/10.1016/j.system.2013.07.013
  23. Rodgers, M. P. H., & Webb, S. (2011). Narrow viewing: The vocabulary in related television programs. TESOL Quarterly, 45(4), 689-717. https://doi.org/10.5054/tq.2011.268062