FSRS vs SM-2: Which Spaced Repetition Algorithm Is Better for Mandarin?

If you use Anki, Mochi, RemNote, or any serious flashcard app for Mandarin, you will eventually meet two acronyms: SM-2 and FSRS. SM-2 is the algorithm Anki shipped with for almost two decades. FSRS is the open-source successor that Anki now bundles as the modern default. For Mandarin learners juggling 3,000+ characters, the difference is not academic. It is roughly 20 to 30 percent fewer daily reviews for the same recall rate.

This article is a head-to-head comparison. It covers what each algorithm actually does, what the published research says about retention and review load, and how to pick the right one for a deck full of hanzi (汉字), pinyin (hànyǔ pīnyīn), and example sentences.

A quick refresher: what is spaced repetition

Spaced repetition is the practice of reviewing a fact at increasing intervals, timed to land just before you would otherwise forget it. Review too early and you waste time. Review too late and you have to relearn the card from scratch. The trick is to predict that “about to forget” moment thousands of times across thousands of cards, each with its own difficulty and history.

Every spaced repetition algorithm answers three questions:

1. How likely are you to recall this card right now?
2. How should that estimate change after every review?
3. When should the next review happen so the recall probability lands near your target (usually 90 percent)?

SM-2, FSRS, and the older Leitner box system are three different answers to those three questions. FSRS is the most recent and, by every published benchmark, the most accurate.

SM-2: the 1987 algorithm that defined an era

SM-2 was designed by Piotr Wozniak, a Polish-Canadian student who wanted to formalise the intuitions behind his SuperMemo software. The algorithm was first published in 1985 and refined through 1987. It quickly became the reference spaced repetition algorithm, and Anki adopted a modified version of it when Damien Elmes launched the app in 2007.

SM-2 is built from three simple ingredients:

• An “easiness factor” per card, starting at 2.5.
• A four-button rating (Again, Hard, Good, Easy) that nudges the easiness factor up or down on every review.
• A simple interval rule: the next interval is roughly the previous interval multiplied by the easiness factor, with minimum and maximum caps to keep the schedule sane.

When you press Again, the easiness drops and the interval shrinks aggressively. When you press Easy, both grow. The algorithm is elegant, easy to implement, and surprisingly effective for a 1987 design.

The wider reason SM-2 lasted so long is that it was the only serious algorithm that was simple enough to implement, published openly, and worked well enough at scale to power a community of millions. For twenty years, SM-2 was the de facto standard for digital flashcards. The 2020s brought a credible replacement.

What SM-2 gets wrong

SM-2 was a remarkable achievement for 1987, but the algorithm has well-known limitations, documented over decades of community use and now confirmed by formal benchmarks. The four that matter most for Mandarin learners are:

1. The easiness factor is a single number per card

SM-2 stores one floating point value per card and multiplies intervals by it. This conflates two different quantities: how inherently hard the card is (a rare character like 艘, sōu, is harder than a common one like 的, de) and how recently you got it wrong. A single multiplier cannot represent both cleanly, so SM-2 under-reviews hard cards and over-reviews easy ones. For Mandarin, where the top 500 words dominate any text, this gap is expensive.

2. No personal calibration

SM-2’s intervals were calibrated on Wozniak’s own review history in the late 1980s. Your memory is not Wozniak’s memory. SM-2 has no mechanism to fit itself to your data; you can change the global “starting ease” or “easy bonus” in Anki’s deck options, but those are hand-tuned constants, not learned parameters.

3. Poor handling of delayed reviews

If you go on holiday for ten days and return to 500 overdue reviews, SM-2 simply ramps the interval back up after each successful review. It does not recalculate the probability that you have actually forgotten each card based on the time that has passed.

4. No explicit retention target

SM-2 has no concept of “I want to remember 90 percent of these cards on the day they are due.” You set interval caps and minimums by hand, which makes the trade-off between review load and recall hard to reason about.

These issues are not fatal. SM-2 is good enough that millions of people use it every day. They are, however, the reason the open-source community started looking for something better.

FSRS: a modern, open-source, free replacement

FSRS stands for the Free Spaced Repetition Scheduler. It was created by Jarrett Ye, then an undergraduate, who published the formal algorithm in a 2022 paper and shipped a working Anki add-on the same year. The project is now maintained under the open-spaced-repetition GitHub organisation, with ports in Rust, TypeScript, Go, Python, Dart, and several other languages. The core library is MIT licensed.

The headline result of FSRS, validated against real review logs from hundreds of thousands of Anki users, is that for the same target retention rate, FSRS users do roughly 20 to 30 percent fewer reviews per day than SM-2 users. The exact saving depends on your deck and history, but it is consistent across the published A/B benchmarks.

The reason FSRS does more with fewer reviews is that it actually models the shape of the forgetting curve, instead of pretending memory is a simple multiplier.

How FSRS represents memory

Every card in FSRS has three numbers:

These three values form the card’s “memory state.” When you press Again, stability collapses and difficulty climbs. When you press Easy, stability grows fast. R is recomputed every time you open the app and decays smoothly toward zero as the next review approaches.

The algorithm ships with 21 default parameters trained on tens of millions of reviews shared by Anki users who opted in. After your first few weeks of review history, FSRS switches to a model fitted to your history using maximum likelihood estimation. The algorithm literally learns your memory. This is the part that SM-2 could never do.

FSRS vs SM-2: side-by-side

The table below summarises the properties that matter most for Mandarin learners. The numbers in the “reviews to reach 90 percent recall” row are rough rules of thumb from the published FSRS benchmarks; individual results vary by deck and user.

For Mandarin vocabulary specifically, the key row is the last one. A reasonable HSK 5 (Hanyu Shuiping Kaoshi, the standard Chinese proficiency test, level 5) target is around 2,500 words, plus another 1,000 for fluency. A scheduler that wastes 25 percent of your reviews on easy common characters like 我 (wǒ) is genuinely expensive in cumulative time.

What the published research actually shows

The FSRS team’s published benchmarks are the closest thing to neutral evidence in this space. The headline numbers, drawn from the open-spaced-repetition project’s public datasets:

• Lower prediction error. FSRS reduces RMSE (root mean square error) on retention prediction by roughly 5 to 6 percent compared to SM-2 across a range of real-world datasets. The bigger headline number is the review-load saving (next bullet).
• Fewer reviews at the same retention. FSRS users reach their target retention rate with roughly 20 to 30 percent fewer reviews per day than SM-2 users. The saving compounds across years of daily study.
• Better handling of irregular study habits. FSRS’s R value is recomputed from elapsed time on every review. Users who study irregularly or take breaks see a smaller retention penalty than under SM-2.
• Stronger results for vocabulary-heavy decks. The improvement is largest on decks with many similar cards (vocabulary lists, single characters) and smallest on cards that are mostly unique (image occlusion, cloze deletions from prose).

There are caveats. The published benchmarks are dominated by Anki users, mostly studying medical, language, and law content in English. Mandarin-specific benchmarks are smaller in number but show the same direction of effect, with one recent community run reporting roughly a 27 percent reduction in reviews for a 4,000-card HSK deck after switching from SM-2 to FSRS-4.9 with optimised parameters.

The honest summary: FSRS is measurably better than SM-2 for vocabulary-heavy language learning, and the gap grows as your deck grows. The improvement is not magical, but it is real and it is the result of a fundamentally better memory model.

How to switch to FSRS in Anki

If you are already using Anki with SM-2 and want to migrate, the process is low risk:

1. Upgrade to Anki 23.10 or later. FSRS has been a built-in option since November 2023, and Anki 24+ ships FSRS-5 support.
2. Open Deck Options, then Scheduling. Change the algorithm from SM-2 to FSRS.
3. Wait until you have at least 1,000 review logs, then run Optimise. FSRS fits its parameters to your specific history.
4. Pick a target retention. 0.85 to 0.90 is a sensible default for Mandarin vocabulary. Lower if you want fewer daily reviews, higher if you want maximum recall.
5. Keep studying. The model gets sharper with every review.

You do not need to reset your deck. The historical review history is preserved and FSRS uses it to fit its initial parameters. Intervals will look different on the first day, but the underlying memory state carries over. Most users see a noticeable drop in daily review volume within two weeks, once the optimisation pass has enough data.

For a deeper walk-through, the FSRS spaced repetition hub collects benchmarks, migration guides, and version-by-version release notes.

How ReadDeck uses FSRS by default

ReadDeck is the offline Mandarin reader we ship. The flashcard surface inside the reader schedules every word you mark as Unknown, including example sentences pulled from the lesson you just read, using an embedded FSRS scheduler. You do not have to configure anything. There is no algorithm dropdown, no retention target slider, no “Optimise” button to forget about. FSRS is on by default.

Why FSRS specifically, and not SM-2 or a hand-rolled interval:

• Mandarin is a high-card-count language. HSK 5 plus 1,000 extra words is roughly 3,500 cards. A scheduler that wastes 25 percent of reviews on 我, 的, and 是 is genuinely expensive.
• The deck shape is lopsided. Frequency drops steeply after the top 500 words. You see 的一千 times more often than 艘. SM-2 treats them identically. FSRS does not.
• Mobile review sessions have to be short. FSRS’s daily-review prediction lets ReadDeck surface a small, high-value batch each morning instead of an overwhelming backlog.
• Offline-first matters. Because FSRS is a small Rust core, we can embed it in the binary. No scheduler call leaves the device, no account is required, and your review history stays on your phone.

The audio companion listens for the same words when you read aloud, and the same memory state drives both surfaces. Words you struggle to pronounce, or that the speech recogniser flags as uncertain, get scheduled sooner, in line with FSRS’s retrievability update.

If you are choosing a Mandarin learning tool in 2026, picking one that has already made this decision for you is the simplest way to benefit from the research. ReadDeck has. SM-2 is still the default in older Anki builds and in apps that have not updated. The research is clear about which one to pick.

FAQ

Is FSRS really better than SM-2?

Yes, by every published benchmark. The open-spaced-repetition team’s A/B tests on real review logs show FSRS reaching the same target retention with roughly 20 to 30 percent fewer reviews per day. The improvement is largest on vocabulary-heavy decks, which is exactly the shape of Mandarin.

Will I lose my review history if I switch from SM-2 to FSRS?

No. Both Anki and ReadDeck preserve your review history across the switch. FSRS uses the first few hundred logs to fit its initial parameters, then continues to refine them as you study. You will see different intervals on day one, but the cards themselves do not reset.

When did FSRS replace SM-2 in Anki?

FSRS became a built-in option in Anki 23.10 (November 2023). It was opt-in at first, then promoted toward the default in subsequent releases. As of 2026, SM-2 is still available as a legacy option for users who prefer it, but FSRS is what new users see.

Does FSRS work for tonal languages like Mandarin?

Yes, and arguably better than for non-tonal languages. Mandarin vocabulary has long, lopsided frequency curves, lots of similar characters, and many cards that compete for the same memory slot. FSRS’s per-card D, S, R state handles this shape better than SM-2’s single easiness factor.

What is the best target retention for Mandarin?

For most adult learners, 0.85 to 0.90 is a sensible default. If you want fewer daily reviews and can tolerate occasional lapses, drop to 0.80. If you are preparing for HSK and want maximum recall, push to 0.90 or 0.92. FSRS makes the trade-off explicit; SM-2 hides it in the formula. To see how your current retention stacks up against an actual exam, HSK Quiz runs free adaptive mocks that grade you against the official HSK pass mark.

Try FSRS without the spreadsheets

Configuring FSRS in a generic flashcard app is fine, but it is not what most Mandarin learners actually want to do with their evening. ReadDeck embeds FSRS inside a Mandarin-first reader, so every paragraph you read, every sentence you listen to, and every card you review shares one scheduler that knows your real memory.

Open ReadDeck and let FSRS decide what you should review tomorrow.