HSK Indonesia Report 2026: The Anatomy of Indonesian Learners' Errors and the Structure of Market Demand

1. The cognitive-demand gradient: errors follow what the question asks, not how long the text is

Each HSK part demands a different cognitive operation. We classified the 34 parts of the HSK 1–5 listening and reading papers into four demand tiers, following the reading-process model of Khalifa and Weir (2009), which separates form recognition, meaning retrieval and syntactic integration: recognition (true/false, picture matching), retrieval (choosing the answer to a short dialogue, matching questions to answers), comprehension (long discourse), and integration (cloze, 选词填空 xuǎn cí tiánkòng, filling a sentence gap from a word bank; and sentence ordering, 排列顺序 páiliè shùnxù, arranging sentence fragments).

The result is a clean gradient. Pooled across levels, recognition items are wrong 13.7% of the time, comprehension 18.3%, retrieval 20.4% and integration 26.2%, or 1.9 times recognition. The surprise is not integration at the top but comprehension below retrieval: long passages are not harder than short dialogues, because a passage supplies context that compensates for unknown words.

The gradient is stable across levels. Integration is the hardest tier at every level from HSK 1 (24.4%) to HSK 5 (38.3%). Recognition moves from 12.6% to 19.1% over the same range. The distance between recognising a word and assembling it into a sentence does not close with level; it widens.

The split by exam section supports this reading. At every level from HSK 1 to 5, the reading section produces more wrong answers than the listening section (HSK 2: 22.1% versus 11.7%; HSK 5: 28.7% versus 19.1%), and almost all integration items sit in the reading section. Across 30 full-paper simulations the two section scores correlate strongly (r = 0.73): a learner weak in one is generally weak in the other, so the section gap reflects question demand, not two different learner populations.

2. Error concentration: a few words and formats produce most of the errors

Pareto analysis answers a practical question: how concentrated are the errors, and therefore how much return can a small intervention yield? We measure it by the share of wrong answers produced by the top 20% and 30% of units, and by the Gini coefficient of error rates across units (0 = even, 1 = fully concentrated).

At the word level, concentration is strong. Among 126 HSK 1 words with at least 20 attempts, the top 20% of words produce 65.4% of all wrong answers, and the top 30% produce 76.1%. Gini 0.52. At the exam-part level, concentration is much weaker: the top 20% of parts produce 44.8% of errors, Gini 0.22.

The difference matters strategically. Vocabulary difficulty can be resolved with a short, targeted list; format difficulty cannot, because it is spread across many parts and must be addressed through general skill. The ten hardest parts below show the spread: five levels, two sections and four different formats appear in the list.

3. Recognition and recall are two different difficulties

Multiple-choice exercises measure recognition: the learner sees the word and picks its meaning. Spaced-repetition flashcards measure recall: the learner must retrieve the meaning from memory and marks the card as forgotten on failure. If both measures captured the same difficulty, words often wrong in exercises should also be often forgotten in flashcards.

They are not. Across 104 HSK 1 words with enough data on both measures, the Spearman correlation between exercise error rate and flashcard lapse rate is only 0.26, and 19.2% of words sit in the top quartile on one measure but the bottom half on the other. Corpus frequency, often assumed to determine difficulty, does not correlate with error rate at all (0.02).

The pattern by word class explains the split. In exercises, function words (particles, adverbs, question words, modal verbs) are the class with the highest error rate, 5.4% (95% CI 4.7–6.3), above content words at 4.7% (4.1–5.3). In flashcards the order reverses: measure words are forgotten most often, 10.3% of cards (8.6–12.3), then content words at 8.1%, while function words sit at only 5.6%.

The explanation is cognitive. Function words have no concrete referent; their meaning depends on position in the sentence, so they are hard to pick in a multiple-choice item that presents them out of context, yet easy to remember because they occur in almost every sentence. Measure words and concrete nouns are the reverse: easy to pick when present, but their characters rarely recur in other sentences, so they fade quickly. Indonesian, which has no question particle comparable to 吗 (ma) and only a handful of classifiers, amplifies both difficulties.

4. Learnability: the hardest format improves fastest when repeated

Hard and untrainable are different things. To measure learnability, we compared a learner's error rate on their first attempt at a level with their repeat attempts at the same level, by cognitive-demand tier. Comparing within the same level keeps the change from being confounded with rising material difficulty.

Integration items, the hardest tier, improve the most: from 31.6% wrong on the first attempt to 25.3% on repeats, a relative fall of 20%. Comprehension falls 7%. Recognition and retrieval do not improve at all (12.8% to 14.1%; 19.8% to 21.3%).

This reading is consistent with Finding 3: integration errors depend on procedural skill (recognising sentence patterns and collocations) that forms through repeated exposure, whereas the recognition errors that remain after one attempt are pure vocabulary errors, which repeating a paper does not cure.

As a methodological note, the same comparison on chapter exercises cannot be read as learnability, because later attempts occur in harder chapters. In chapter exercises the error rate rises with attempt order (word arrangement from 12.2% to 22.7%), which is a difficulty ramp, not a decline in ability.

5. The level wall: exam difficulty is not linear, and the variance widens

Simulation pass rates (a score of 60% or above) map exam difficulty from the Indonesian learner's point of view. The map is not linear. HSK 1 and HSK 2 pass at 81% and 80%. HSK 3 drops to 66%, HSK 4 recovers to 78%, and HSK 5 falls to 41% with a median score of 42 out of 100.

The first dip at HSK 3 coincides with the removal of pinyin from the paper and the arrival of the writing section; learners who lean on pinyin lose their support here. The recovery at HSK 4 reflects a filtered population: those who reach HSK 4 have generally crossed the character transition. The HSK 5 wall is different in kind: a jump in vocabulary and text length that exam strategy can no longer bridge.

More informative than the median is the spread. The interquartile range widens from 35 points at HSK 1 (65–100) to 62 points at HSK 3 (32–94) and 69 points at HSK 5 (13–82). At higher levels learners split into two populations: the ready and the premature. An average pass rate hides this polarisation.

6. The structure of demand: volume sits with beginners, willingness to pay with HSK 3–4

The final section describes the market rather than the learner. The question: at what level do Indonesians decide to pay for a Mandarin learning product, and how does that level compare with the distribution of the learner population? All figures are percentages; absolute counts of paid users are not published.

We use two measures. First, a representation index: the share of paid users with a given placement-test result divided by the share of all placement-test takers with that result. An index of 1 is proportional; above 1 means the group is more inclined to pay. Second, a conversion rate: the share of takers with a given result who subsequently became paid users.

Both measures agree. Takers placed at HSK 3 have an index of 2.16 and a conversion of 33%; HSK 4, an index of 1.61 and 25%. Takers placed at HSK 1, who are 48% of all takers, have an index of 0.79 and a conversion of 12%. Takers placed at HSK 5–6 have indices below 0.45: they already have their own way of studying, or their needs exceed a structured product.

In volume terms, however, beginners still dominate: 38% of paid users placed at HSK 1, and by highest chapter studied 41% are still at HSK 1. Seven in ten paid users choose the full HSK 2–6 bundle, and among those placed at HSK 3 the share is 80%. Buyers do not buy a level; they buy a path to a destination.

For the industry these are two segments with different economics. The beginner segment is large but converts poorly, so it demands cheap acquisition and strong onboarding. The HSK 3–4 segment is small but converts at 2.5 times the rate and prefers the full bundle, so it tolerates higher prices and deeper products. With roughly 5,000 candidates per official HSK sitting in Indonesia in 2024 (two media-reported sittings: 4,970 and 5,547 candidates), the addressable market for HSK 3–4 products is in the thousands per year, not tens of thousands, but at a far higher value per learner.

Implications

Methodology

Exam simulations

439 simulations completed by 140 Indonesian learners on Pinter Mandarin using official HSK papers (HSK 2.0 format) at levels 1–6, 9,689 answers in total. Only the listening and reading sections are auto-scored; writing is excluded. Error rates are computed on answered questions; skipped questions are not recorded, so true rates are equal or higher.

Cognitive-demand tiers

The 34 parts of the HSK 1–5 papers were classified manually into four tiers (recognition 12 parts, retrieval 8, comprehension 8, integration 6) by the operation the question demands, following the reading-process model of Khalifa and Weir (2009).

Error concentration

Share of wrong answers from the top 20% and 30% of units (ranked by number of wrong answers), and the Gini coefficient of error rates across units. Units = exam parts (34) and HSK 1 words with at least 20 attempts (126).

Word classes and confidence intervals

Word classes follow the part of speech in the HSK 3.0 list: function words (particles, adverbs, pronouns and question words, prepositions, conjunctions, modal verbs), content words (nouns, verbs, adjectives, numerals), measure words, set phrases and greetings. 95% confidence intervals use the Wilson method.

Recognition vs recall

Spearman correlation between exercise error rate and flashcard lapses per card across 104 HSK 1 words with at least 30 attempts and 30 cards. Correlation of corpus frequency with error rate on the same words.

Learnability

Per learner-level pair (91 pairs), error rate on the first attempt compared with repeat attempts at the same level, by demand tier. The same comparison on chapter exercises is reported as a difficulty ramp, not learnability.

Level wall

Score = total score divided by maximum score per simulation (full-paper or single-section mode), writing-only simulations excluded; pass = 60% or above. Medians and quartiles reported.

Market

Paid user = successful transaction or active paid access; only percentages are published. Representation index = share of paid users per placement result divided by share of all takers with that result. Official 2024 candidate counts are cited from media reports (see references).

Exclusions and period

Internal accounts excluded. HSK 6 excluded from claims (few simulations). No individual data processed beyond aggregation. Data as of 12 September 2026, activity June–September 2026. Refreshed annually.

Download data: laporan-2026.json (CC BY 4.0)

Statement and assumptions

This report rests on the following assumptions. Those who cite its figures are advised to carry the relevant assumption with the figure.

1. Learners who take simulations on this platform represent Indonesians studying independently or alongside a course; they do not necessarily represent all official HSK candidates.
2. The papers used follow the HSK 2.0 format. The HSK 3.0 standard tested from 2026 uses a different section structure; the part ranking must be recomputed once 3.0 data are available, although the cognitive-demand gradient is expected to hold.
3. The classification of demand tiers and word classes was done manually and is debatable at the margins; the classification code is open to review.
4. Error rates are computed on answered questions. Questions skipped when time ran out are not recorded, so the final parts of each session are likely harder than they appear.
5. Paid-user figures are percentages of this platform's paying population and cannot be generalised as a national market measure. Representation indices and conversion rates for the HSK 2, 5 and 6 groups come from small groups and are indicative.
6. Correlations and comparisons are descriptive; no significance testing was performed beyond the confidence intervals in the word-class exhibit.

References and aligned reading

We set our findings beside the available research. Each reference below was read directly; the sentence under each explains how it relates to our finding. Three gaps are worth stating: no study compares reading and listening error rates for Indonesian learners, none separates recognition from recall by word class at HSK 1, and no official level distribution of Indonesian HSK candidates is published. On those three points this report presents new data.

1. 1.

   Khalifa, H. & Weir, C. J. (2009). Examining Reading: Research and Practice in Assessing Second Language Reading (Studies in Language Testing 29). Cambridge University Press / Cambridge ESOL. https://www.cambridgeenglish.org/research-and-validation/published-research/studies-in-language-testing/

   The layered reading-process model (word recognition, meaning retrieval, syntactic integration, discourse) that underlies the four cognitive-demand tiers in Exhibit 1.

2. 2.

   People's Daily Online (人民网) (2024). 印尼举行汉语水平考试，4970名考生在23个考点参加 (Indonesia holds HSK: 4,970 candidates at 23 test centres, March 2024). world.people.com.cn, laporan berita. http://world.people.com.cn/n1/2024/0318/c1002-40197824.html

   Source of the March 2024 official-sitting candidate count used to size the market in Finding 6. A media report, not official per-level statistics.

3. 3.

   Harian Inhua (2024). Ujian HSK Oktober 2024 di Indonesia: 5.547 peserta di 18 lokasi. koran.harianinhuaonline.com, laporan berita. https://koran.harianinhuaonline.com/2024/10/22/106047/

   Source of the October 2024 official-sitting candidate count (Finding 6).

4. 4.

   Irawati, R. P. & Anggraeni (2018). Analisis kesulitan mahasiswa dalam memahami teks 阅读 pada HSK level IV. Longda Xiaokan 1(2), Universitas Negeri Semarang. https://doi.org/10.15294/longdaxiaokan.v1i2.12689

   Among 24 Mandarin-education students, the hardest HSK 4 reading parts were the fill-in-the-blank items and arranging three sentences into a paragraph. Directly aligned with Findings 1 and 2.

5. 5.

   Zhou, J. (2022). The effects of syntactic awareness to L2 Chinese passage-level reading comprehension. Frontiers in Psychology 12:783827. https://doi.org/10.3389/fpsyg.2021.783827

   Among 209 L2 Chinese learners, word-order knowledge predicted cloze scores more strongly than multiple-choice scores. Explains why integration items punish grammar weakness (Finding 1).

6. 6.

   Cai, J., Han, Y. & Jiang, X. (2025). ClozCHI: A cloze test for measuring L2 Chinese proficiency from novice to advanced levels. Behavior Research Methods. https://doi.org/10.3758/s13428-025-02834-9

   Among 225 HSK 3–6 candidates, cloze scores track HSK level closely; cloze is a level-discriminating format. Consistent with its position at the top of the cognitive-demand gradient (Finding 1).

7. 7.

   Zhou, Q., Du, F., Lu, Y., Wang, H., Herman & Yang, S. (2024). The development of reading comprehension ability of Chinese Heritage Language learners in Indonesia. Language Testing in Asia 14. https://doi.org/10.1186/s40468-024-00276-2

   Among 275 Indonesian learners, reading sub-skills (vocabulary, syntax, discourse) mature only in adulthood. Supports reading as the lagging skill (Finding 1), though it does not compare it with listening directly.

8. 8.

   Ariqo, U. M. (2022). Analisis kesalahan penggunaan partikel modal 吗, 呢, 吧. Skripsi, Universitas Hasanuddin. https://repository.unhas.ac.id/id/eprint/13476/

   A 54.9% error rate on modal particles, with wrong-function selection dominant. Aligned with the high error rate on 吗 (ma, question particle) within the function-word class in Finding 3.

9. 9.

   Arlim, G. A. (2024). 印尼汉语学习者新HSK1-5级易混淆词语分析. Journal of Maobi 2(1), Universitas Sebelas Maret. https://doi.org/10.20961/maobi.v2i1.87617

   Indonesian learners' HSK 1–5 confusions cluster in homophones and near-synonyms, driven by L1 transfer and collocation differences. Supports Findings 2 and 3.

10. 10.

    Trihardini, A. (2015). Kesalahan penggunaan kata bantu bilangan bahasa Mandarin pada siswa Indonesia tingkat prapemula. Lingua Cultura 9(1), 7–12. https://doi.org/10.21512/lc.v9i1.755

    Beginner learners systematically omit or mis-select classifiers. Aligned with 件 (jiàn) as the most-forgotten word (Finding 3).

11. 11.

    Veronica, T., Yankhiong, B., Thamrin, L., Suhardi & Lusi (2023). Pengaruh bahasa Khek terhadap penguasaan kata bantu bilangan Mandarin. Edukatif 5(6). https://doi.org/10.31004/edukatif.v5i6.5831

    Classifier errors concentrate among beginners, with Hakka and Indonesian interference cited. Supports Finding 3.

12. 12.

    Qin, W. (2020). Penyebab-penyebab kesalahan penggunaan kata bahasa Mandarin: tinjauan terhadap mahasiswa jurusan bahasa Mandarin di Indonesia. Linguistika 27(2), Universitas Udayana. https://doi.org/10.24843/ling.2020.v27.i02.p01

    Among 44 students, Indonesian interference is the main linguistic cause of word-usage errors. Background for Finding 3.

13. 13.

    Go, Y. (2014). Error analysis of Chinese word order of Indonesian students. Humaniora 5(2), Universitas Bina Nusantara. https://doi.org/10.21512/humaniora.v5i2.3250

    Adverbial (time, place, preposition) ordering is the recurrent Indonesian error. Relevant to sentence-ordering items and word-arrangement exercises (Findings 1 and 4).

14. 14.

    Wang, H. 王红侠 (2017). 印尼学生汉语习得的偏误类型和成因. 海外华文教育 2017(1). https://m.fx361.com/news/2017/0310/18571650.html

    Beginner Indonesian errors split into syntax, lexis and expression, mostly from L1 analogy. Background for Findings 1 and 3.

15. 15.

    Junaeny, A. & Adam, M. R. (2022). An error analysis in Chinese language tones pronunciation by Indonesian students. IJSASCS 6(1), Universitas Sebelas Maret. https://doi.org/10.20961/ijsascs.v6i1.70790

    Tone errors rise sharply in disyllabic words; the non-tonal L1 is the main cause. Explains why listening feels hard early on, before our data show reading as the persistent weakness (Finding 1).

16. 16.

    Yi Ying, Suprayogi, M. N. & Hurriyati, E. A. (2013). Motivasi belajar bahasa Mandarin sebagai bahasa kedua. Humaniora 4(2), Universitas Bina Nusantara. https://doi.org/10.21512/humaniora.v4i2.3579

    Among 276 university students, motivation was low with no integrative/instrumental gap. No published study covers paying behaviour by HSK level; Finding 6 is new data.

Related articles on Pinter Mandarin

• HSK test guide: dates, fees, venues and registration (Indonesian)

  Context for the exam analysed in this report.

• HSK 1: complete guide to format, vocabulary and passing (Indonesian)

  The HSK 1 section formats discussed in Exhibit 3.

• The 100 most-used Mandarin verbs (Indonesian)

  Modal verbs such as 想 (xiǎng, want) with example sentences.

• Basic Mandarin grammar (Indonesian)

  The sentence positions of 吗 (ma), 太 (tài) and 哪 (nǎ), the error source in Finding 3.

• The four Mandarin tones (Indonesian)

  Why listening feels hard at first, and why the data show reading as the persistent weakness.

• HSK simulations on official past papers

  Practise the integration items (cloze, sentence ordering) that improve 20% on a repeat attempt.

How to cite

Pinter Mandarin (2026). HSK Indonesia Report 2026: The Anatomy of Indonesian Learners' Errors and the Structure of Market Demand. https://www.pintermandarin.com/en/research/hsk-indonesia-report-2026. Licensed CC BY 4.0: free to quote and share with attribution.