How hard is each HSK level

HSK 1 words account for about 60% of the words in the subtitles we measured. You can know them all and still miss what a sentence means.

If you're studying HSK and want something to read alongside it, Dengdeng helps you read real stories from China at your level. You can use the vocabulary you know and look up the words you still need.

The early HSK words come up again and again. Later bands add fewer word occurrences to the count, but those less frequent words can be essential to understanding a sentence.

I measured all 10,937 words in HSK 3.0 against a frequency list from 33.5 million words of film and TV subtitles, SUBTLEX-CH.

The new standard folds 7, 8 and 9 into a single top list. We use subtitles as a proxy for everyday speech. Here, "coverage" means the share of word occurrences in that corpus found on an HSK list. Every repetition counts.

These numbers describe vocabulary coverage in subtitles. They do not measure how much of a conversation you would understand.

What "coverage" here means

• Subtitles are speech, not writing. SUBTLEX is dialogue. A word that's common in newspapers but rare in conversation — 该 (the formal "should"), 及 (the written "and"), much of the HSK 7-9 vocabulary — will look low-coverage here even though you still need it in writing. Read every number as "in everyday speech," and add a mental asterisk for formal registers.
• HSK 3.0 is not a frequency list. The committee chose words for a curriculum, not by raw count, so the bands don't line up perfectly with frequency. That mismatch is exactly what we're measuring.
• Frequency isn't usefulness. Recognising a word isn't understanding a sentence, and "covered" tokens include the easy glue words. A coverage percentage does not give you a minimum level of understanding.
• A few points of coverage can still decide a sentence. Later bands add a few percent of spoken tokens. Those words are often the ones that reverse who did what, whether something happened, or which way a claim went. Missing them does not leave a vaguer version of the same idea. You keep the glue and lose the point.
• Tagging is automatic. Part of speech comes from jieba at the dictionary level, per word, not disambiguated in context. It's approximate, and we only lean on it for the broad shape.

HSK 1 word coverage

The 497 words in HSK 1 account for 60.5% of word occurrences in the corpus. About three words in five, counting repetitions. The other six bands, all 10,440 words of them, add another thirty percentage points between them.

The missing two words in five may tell you what happened, to whom, or why. If those are the words you do not know, the whole sentence can remain unclear. Knowing most of its words is no guarantee that you get even the main idea.

That 60.5% is an average across the corpus. Individual sentences can have very different coverage.

The curve climbs fast and then flattens hard. Finish HSK 2 and you're at 72.1%. HSK 3 takes you to 78.8%, HSK 4 to 82.1%, HSK 5 to 84.8%, HSK 6 to 86.8%. Take the entire 7-9 band — 5,597 more words, more than the first six levels put together — and you reach 90.5% and no further.

About a tenth of the word occurrences in this corpus fall outside the HSK lists.

This is not a flaw in HSK. A small set of words does most of the talking. Finishing a level early is worth a lot. Grinding the top bands to "completion" buys you single-digit percentages.

How often those words appear is not how much they can change a sentence.

Coverage added by each level

Stack up how much each band's new words add to the coverage count.

HSK 1 buys +60.5 points. HSK 2 buys +11.6. HSK 3, +6.7. By HSK 4 you're paying 972 new words for +3.3 points. HSK 5 buys +2.7, HSK 6 a mere +2.0. The 7-9 band looks like a rebound at +3.8 — but that's 5,597 words doing the work, the worst words-per-point ratio on the chart by a wide margin.

By HSK 3, fewer than 2,200 words cover over 78% of word occurrences. After that, each band adds roughly two to four percentage points, despite containing hundreds or thousands of new words.

You still learn later words to understand what people say. They also let you read harder texts, handle formal registers, and pass exams. Their smaller share of the count does not tell you how much you need them in a particular sentence.

Word types

Coverage is only half of "hard." The other half is the kind of word. Tag every band's new words by part of speech and the mix of word types changes as you go up.

In HSK 1, function words — pronouns, particles, prepositions, measure words, adverbs — are 21% of the new vocabulary. By HSK 3 they are down to 12%, and by 7-9 just 5%. Each word is counted only in the first band where it appears.

That is a pattern in the vocabulary lists, not a measure of how much grammar you have learned. Knowing a particle does not mean you know every sentence pattern that uses it.

What comes next is content. Nouns are the largest slice throughout, holding steady around 44-49% in every band. The slice that grows as you climb is verbs: 25% of HSK 1's new words, then 33% by HSK 4 and a peak of 37% in HSK 5.

Verbs make up a larger share of the middle bands. You keep adding nouns too, while learning how familiar words work in less familiar sentences.

Later bands are rarer

If the early levels are easy because their words are everywhere, the later levels are hard for the mirror reason: their words are rare. Take the words in each band that appear in SUBTLEX at all, and look at where they sit in the frequency list.

HSK 1's median word is the 645th most common in the language — you'll see it constantly. HSK 2's is #1,676, HSK 3's #2,534. Then the climb steepens: HSK 4 sits at #4,049, HSK 5 at #5,212, HSK 6 at #6,665, and the 7-9 median word is the 12,133rd most frequent — a word you might go a whole conversation without hearing.

A top-band word isn't harder because the characters are stranger. It's harder because you have fewer chances to see it in use.

Across every band, 97-99% of the HSK list appears in the subtitle corpus at all — only the 7-9 band dips, to 97%. HSK 3.0 is not padded with unused textbook words. These are words people say. The upper ones are just rare. The handful that don't appear are mostly written-register or specialist terms — the spoken-corpus caveat predicts those would be missing.

Each level

The four signals together:

• HSK 1: 497 words account for about 60% of word occurrences in the subtitle corpus. The words you do not know can still leave a whole sentence unclear.
• HSK 2 and 3 bring coverage from about 60% to 79%. Later bands keep adding coverage, but in smaller increments.
• HSK 4 through 6 are a different kind of work. Coverage creeps up two or three points a level; the words are verbs and nouns ranked in the thousands. Those points still buy range, formality, and the precise words that decide a sentence.
• HSK 7-9 is not a final rung. Five thousand-plus rare words for the last few points of spoken coverage — and where the formal, written register that subtitles can't see finally comes into its own.

The upper levels are a different project from the lower ones, and you start them for different reasons.

You can start a short supported text alongside HSK study. Try a paragraph with word lookup and check whether you can follow what happens; the coverage figure alone cannot tell you. Chinese graded readers by HSK level lists what fits each band.

Common questions

How many words are in HSK 1?

497 distinct words in the HSK 3.0 list we measured. They account for about 60% of word occurrences in the subtitle corpus, not 60% of its meaning.

How many words are in HSK 3.0?

10,937 distinct words. Levels 7, 8 and 9 share one list of 5,597.

Do I need to finish a level before I start reading?

No. Try a short text with word lookup and a translation you can check. Choose the difficulty by whether you can follow the paragraph, rather than by a corpus coverage percentage.

Are later HSK levels worth it if they barely add coverage?

Yes. Coverage counts every glue word. The extra few percent are often the words that decide a sentence.

Method

Word lists: HSK 3.0, 10,937 distinct simplified words after de-duplication; each word is counted toward the lowest band it appears in, so per-band figures describe new vocabulary. Frequency and coverage: SUBTLEX-CH (Cai & Brysbaert 2010), ~33.5M word-segmented subtitle tokens, treated as a proxy for spoken register. Cumulative and marginal coverage are token-weighted sums over each band's words; "in-corpus share" and "median rank" use only the band's words that appear in SUBTLEX (97-99% of each list). Part of speech is from jieba.posseg at the dictionary level (per word, not context-disambiguated); POS charts are type-weighted (share of the word list, not of speech) and are approximate. SUBTLEX is subtitles, so words frequent in writing but rare in speech read low here — the register caveat applies throughout, most of all to HSK 7-9. A native-speaker review of the band-by-band reading is pending.

For reading practice alongside your HSK study, take a look at Dengdeng's daily edition. Choose a level you can follow, then tap the unfamiliar words for pinyin and meaning. One story from each daily edition is free in the app.