Every source.
Its edition, its licence, its limits.

Covers

125,264 source records (header declares 125,264) plus 136 separately attributed zimai editorial word records. 122,686 distinct traditional and 121,500 distinct simplified headwords; the two forms differ in 77,917 records. 13,528 single-character and 111,736 multi-character records; 199,737 definition segments, 1,460 with classifier notes, 20,199 proper names.Mandarin (Pinyin) → English · 繁 and 简

Licence & attribution

CC-CEDICT contributors; published by MDBG. Portions © 1997–98 Paul Andrew Denisowski (CEDICT).

How zimai transforms it

Every source row, numbered reading and definition boundary is kept with its SHA-256 line hash. Display Pinyin is derived; search indexes fold case and tone marks without merging records. The declared ranking (definition-grade v2) orders results; it is not frequency.

Known gaps

No Cantonese readings and no example sentences. Glosses are terse; a record is not a sense inventory. No usage frequency.

Freshness & update path

Edition 2026-10-13, 2 days before 2026-10-15. Upstream publishes a new export almost daily. Update path: re-import the pinned file, rebuild indexes, rerun the ranking script and the relevance benchmark.

Speed

CC-CEDICT search partitions: 347 files, 19.4 MB on disk, largest 78 KB (24 KB gzip); loads one or two per query, after the first keystroke. CC-CEDICT word shards: 1,960 files, 38 MB on disk, largest 23 KB (5 KB gzip); loads for visible results. CC-CEDICT entry records: 4,352 files, 229.1 MB on disk, largest 495 KB (118 KB gzip); loads server-rendered entries.

Search CC-CEDICT ↗Word indexEdition, attribution, licenceLicence notice

Covers

160,731 records with original German definitions; 159,362 distinct traditional and 159,056 distinct simplified headwords (forms differ in 104,413 records). 3,165 “Bsp.” example pairs in 944 records; 120,241 records carry a part-of-speech label. Edit history: 380,013 logged versions by 48 usernames, 2011-05-28 to 2026-10-13.Mandarin (Pinyin) → German · 繁 and 简

Licence & attribution

HanDeDict contributors (per-entry usernames retained), published by Gábor L Ugray; originally supported by Chinesisch-Deutsche Gesellschaft e.V. Hamburg under Dr. Michael Klaus Engel and Jan Hefti.

How zimai transforms it

Active rows are parsed without changing spelling, reading or definition segments. Search-only aliases for case, umlauts, ß/ss and marked or numbered Pinyin. The full export with every contributor history ships unchanged as handedict.u8.gz. No English sense is merged in and no German is generated.

Known gaps

No Cantonese–German dictionary exists among the open sources; zimai says so instead of falling back. German “Bsp.” examples are held but not yet shown as sentences. The upstream README states CC BY-SA 2.0; the export header states 3.0, which zimai follows and documents.

Freshness & update path

Edition 2026-10-14, 1 day before 2026-10-15. The pinned commit is still the newest upstream export (checked 2026-10-15). Exports appear after edits. Update path: re-import, rebuild the 804 partitions, rerun ranking.

Speed

HanDeDict search partitions: 804 files, 30.5 MB on disk, largest 176 KB (30 KB gzip); loads only after Meanings in: Deutsch. HanDeDict entry shards: 4,096 files, 210.2 MB on disk, largest 89 KB (16 KB gzip); loads only for German results.

Search HanDeDict ↗Quelle & LizenzFull export with histories

Covers

34,335 entries with source Jyutping; 29,698 distinct traditional and 29,476 distinct simplified headwords (forms differ in 18,112 entries). 34,276 entries also give Pinyin; 9,940 are single characters. 138 nonstandard readings are kept as written; 8,882 entries are marked by Pleco as adapted from CC-CEDICT.Cantonese (Jyutping) → English · 繁 and 简

Licence & attribution

Copyright © 2015–17 Pleco Inc.

How zimai transforms it

Exact forms, source Jyutping, Pinyin reference, definition segments, inline comments and line hashes retained. Slash-separated Jyutping alternatives are indexed separately; all six tone numbers stay significant and Jyutping never passes through Mandarin normalisation. No generated pronunciations.

Known gaps

CC-Canto was designed as a supplement to CC-CEDICT, so many everyday words that CC-CEDICT already lists have no CC-Canto entry: of the 122,686 CC-CEDICT traditional headwords, 8,965 also have one. See the spot check above. No newer edition has been published since 2017.

Freshness & update path

Edition 2017-02-02, 9.7 years before 2026-10-15. Frozen upstream (last edition 2017-02-02, rechecked 2026-10-15). The same publisher’s separate “Cantonese readings for CC-CEDICT” file (2015-09-23, CC BY-SA 3.0) is the known path to Jyutping for CC-CEDICT words; it is not installed.

Speed

CC-Canto search partitions: 135 files, 4.6 MB on disk, largest 35 KB (13 KB gzip); loads only with Chinese: Cantonese. CC-Canto entry shards: 4,350 files, 79.6 MB on disk, largest 198 KB (38 KB gzip); loads only for Cantonese results.

Search CC-Canto ↗Source & licenceOriginal archive

Covers

8,243 character records: all 8,105 of the 通用规范汉字表 (2013), identified by kTGH (3,500 / 3,000 / 1,605 by level), plus 138 other forms. 80,119 further reference values (dictionary locators, Japanese, Korean, Vietnamese, fanqie, Cangjie and more) on 8,243 characters.Characters · Mandarin and Cantonese readings · English gloss

Licence & attribution

Unicode, Inc.

How zimai transforms it

Values are selected by exact code point from the pinned archive and never copied to a variant. Standard-list position is list membership, not a frequency rank.

Known gaps

348 characters have no kDefinition gloss and 41 no kCantonese reading; those gaps stay visible. kCantonese is a character-level reading for collation, not a word pronunciation.

Freshness & update path

Edition 2025-07-24, 15 months before 2026-10-15. Unicode publishes Unihan every September. Unihan 18.0.0 (2026) is available upstream and not imported; the claimed edition changes only with a real migration and diff.

Speed

Character catalog shards: 65 files, 1.2 MB on disk, largest 20 KB (5 KB gzip); loads Characters view and grid.

Search characters ↗Character indexReference valuesSelected input rows

Covers

8,243 exact descriptions for the catalog; 1,096 characters have region-specific alternatives (G, H, T, J, K, P, V tags kept); 32 are atomic.Written form · all regions

Licence & attribution

Andrew West (魏安) Publisher permits personal and commercial reuse; statement and source header retained..

How zimai transforms it

Descriptions and region tags are kept exactly; zimai reads them aloud in words (“left 氵, right 青”) and uses them for parts: search. Shape is not etymology.

Known gaps

Describes modern shapes only, not historical formation. Incomplete descriptions remain marked rather than guessed.

Freshness & update path

Edition 2025-06-27, 16 months before 2026-10-15. No newer file since 2025-06-27 (checked 2026-10-15). Update path: re-select by code point and diff descriptions.

Search by parts ↗Selected records and notice

Covers

Medians for 9,507 characters; 6,995 of zimai’s 8,243 character pages have a stroke-order player (6,963 repackaged medians plus 32 dossier stroke files). The recogniser knows the same 9,507 characters.Written form · mainland stroke order

Licence & attribution

Shaunak Kishore (Make Me a Hanzi) from Arphic PL KaitiM GB and UKai (© 1999 Arphic Technology); recogniser by Gábor L Ugray after Jordan Kiang’s HanziLookup.

How zimai transforms it

Medians are repackaged one file per character with a notice; points and order unchanged. The recogniser ships unmodified as WebAssembly; zimai only prepares the input strokes and fuses two preparations.

Known gaps

1,248 catalog characters have no stroke data. Stroke order is the mainland (PRC) convention only; Taiwan and Hong Kong orders differ for some components.

Freshness & update path

Make Me a Hanzi’s data last changed in 2018. Update path for missing and Taiwan-standard characters: an Arphic-licensed extension such as animCJK (not installed).

Speed

Stroke-order files: 6,963 files, 7.2 MB on disk, largest 2 KB (953 B gzip); loads one per character, on Play. Handwriting recogniser: 3 files, 972 KB on disk; loads only when the pen pad opens.

Licence (Arphic PL)Recogniser licence (LGPL-3.0)Make Me a Hanzi notice

Covers

7,127 readings on 5,274 exact character matches, with fanqie, categories and identifiers aligned.Middle Chinese (Guangyun-based transcription)

Licence & attribution

Guillaume Jacques and Alexandre Delaporte (CRLAO)

How zimai transforms it

Exact catalog intersection; semicolon-aligned fields split without normalisation; nothing copied across simplified/traditional variants.

Known gaps

Generated automatically from Guangyun and checked on a subset by its authors; errors remain possible. Characters outside the matched set have no reading.

Freshness & update path

Edition 2026-01-13, 9 months before 2026-10-15. The pinned commit is still upstream HEAD (checked 2026-10-15).

ScholarshipNormalised selection

Covers

2,623 characters, each with one attributed “聲” (phonetic) component; source row, comment and reference number kept for every one.Character formation

Licence & attribution

Taichi Kawabata and the CJKVI project

How zimai transforms it

Only explicit single-character 聲 labels with both ends in the catalog. No 亦聲/省聲, no variant chains, no inferred semantic roles.

Known gaps

Traditional analyses are not invariably correct (Hill & List 2019). A phonetic series does not prove a shared word origin.

Freshness & update path

Edition 2018-02-20, 8.7 years before 2026-10-15. No upstream change since 2019.

ScholarshipAdapted dataLicence

Covers

8 complete witnesses (論語、孟子、老子、莊子、荀子、韓非子、孫子、詩經): 279 reading sections, 9,629 passages, 350,752 Han occurrences of 5,184 distinct graphs.Literary Chinese · original witness script

Licence & attribution

Kanseki Repository (Kanripo) contributors; base-edition labels kept in each file.

How zimai transforms it

Text is split into passages without rewriting; quotations never follow the script preference. Teaching notes and excerpt translations are an AI-assisted pilot on top, labelled as such.

Known gaps

Contextual teaching notes cover 172 passages, not all. No full translation is imported.

Freshness & update path

Edition 2017-03-23, 9.6 years before 2026-10-15. All eight pinned commits are still the newest on their edition branches (checked 2026-10-15).

Speed

Classical search chunks: 77 files, 2.5 MB on disk, largest 47 KB (16 KB gzip); loads only in Classical search.

Search Classical texts ↗Read the witnessesWitness sources

Covers

39 sentence pairs (33 Mandarin, 6 Cantonese) linked to 53 word records and 179 characters; 446 candidates were excluded for unresolved attribution.Mandarin and Cantonese → English

Licence & attribution

Individual Tatoeba contributors, credited per sentence and per translation. some sentences CC0; each sentence and translation keeps its own licence.

How zimai transforms it

Original spelling, punctuation and translations preserved; added-by contributor, editors and current owner resolved from the official logs.

Known gaps

A small demonstration set, not a corpus. No audio, no German translations yet.

Freshness & update path

Edition 2026-10-13, 2 days before 2026-10-15. Tatoeba publishes weekly exports. A pinned per-language export with a licence manifest is the update path; it is not yet imported.

Sentence readerAttributed records