Rime Wanxiang: a data-first pinyin scheme for Rime

What Rime Wanxiang solves, and who it is aimed at

Rime itself is an input method engine, not a dictionary. A user installs a front end such as Weasel, then supplies a schema and a word list, and the quality of the candidates depends almost entirely on those two files. Most users either accept a small default dictionary or spend evenings merging lists from several sources. Rime Wanxiang is an attempt to remove that work by shipping a complete scheme: a tone-annotated dictionary, a grammar model, auxiliary code support and a set of Lua scripts, packaged so that the schema file is the entry point.

The project describes its foundation as the Wanxiang dictionary, which it says was optimised with AI and a large corpus and has now entered a manual maintenance phase. The README states that the dictionary carries tone annotation, phrase composition and word frequency as its basis, and that it is built for sentence-level input rather than single-word lookup. That is the pitch: fewer page turns, because the frequency data and grammar model push the intended candidate to the front.

The audience is narrower than the repository topics suggest. Anyone using Rime with full pinyin or double pinyin is the target. Users of non-pinyin layouts appear only through third-party schemes in the ecosystem section, such as Li's triple pinyin and Hu code, which build on the Wanxiang tone dictionary rather than on the main schema. If you use a shape-based or code-based input method, this is not your project.

The four editions and what each one assumes about your machine

The README splits the project into Base, Pro, Lite and Pure, sharing the same data foundation but differing in feature completeness, auxiliary codes, Lua extensions and runtime requirements.

Base is the default recommendation and ships the full tone-annotated dictionary with the complete Lua extension set. Pro targets heavy double pinyin users who want auxiliary code input, word creation and finer filtering, and it carries its own auxiliary-code dictionary rather than the standard one. Lite keeps full pinyin and double pinyin but drops the tone data, prediction and complex word creation, and trims the heavier Lua modules. Pure ships no Lua at all and is aimed at Windows 7, fcitx4-rime, or any environment that cannot run Lua.

That last point is the real dividing line. The feature differences between Base and Lite are matters of taste. The difference between Pure and everything else is a hard runtime constraint. If your Rime build has no Lua interpreter, three of the four editions will not work as documented, and the README says so directly rather than leaving it to be discovered.

The schema files are named per edition: wanxiang.schema.yaml, wanxiang_pro.schema.yaml, wanxiang_lite.schema.yaml and wanxiang_pure.schema.yaml. The repository also contains separate schemas for abbreviation, T9, mixed code, English, reverse lookup and phrase input, which suggests the main editions are not the only entry points, though the README does not walk through those.

How the dictionary, grammar model and Lua layer fit together

The data flow follows standard Rime practice. A schema file declares which dictionaries to load, which translators to run and which filters to apply. Rime reads the .dict.yaml files, builds a binary table, and at input time the translator proposes candidates that the grammar model and frequency data reorder.

What Wanxiang adds to that pipeline is mostly in the data. The dictionary entries carry tone annotation, which the project treats as an input dimension rather than metadata: the README lists tone-annotated pinyin as one of the three foundations of the experience, alongside phrase composition and precise word frequency. A grammar model then scores multi-word sequences so that a plausible sentence beats a string of individually frequent words. This is the mechanism behind the sentence-flow claim, and it is the part that cannot be judged from a README. Whether the frequency data matches your writing is an empirical question about the corpus, and the project does not publish that corpus.

The Lua directory holds the extension layer. The README names calculators, super annotations, symbol wrapping, dynamic timestamps and a symbol library that accepts thousands of Unicode symbols by name. These are optional conveniences rather than part of candidate generation, which is why Lite can drop some of them and Pure can drop all of them without losing the dictionary. Treat the Lua scripts as a separate product bundled in the same repository. If you want the candidates but not the extras, Lite exists precisely for that.

Installing Rime Wanxiang and typing your first sentence

The README does not contain install commands. It links to the project documentation at amzxyz.github.io/rime-wanxiang, where the quick-start page covers deployment on Windows, macOS and iOS or Android. Follow that page for your platform, since the steps depend on your Rime front end.

What the repository does tell you is which file to select after the files are in place. For a first install the README recommends Base, whose main schema is wanxiang.schema.yaml. In most Rime front ends you pick a schema by name from the deployment menu rather than by filename, so the entry you are looking for is the one backed by that file.

After deployment the scheme is active. A first useful check is to type a full sentence rather than a single word, because that is the behaviour the grammar model is meant to affect. The repository root contains the schema and dictionary files that Rime reads:

wanxiang.schema.yaml
wanxiang.dict.yaml
custom_phrase.txt
custom_phrase.dict.yaml

The expected result is a multi-word candidate near the top of the list rather than a sequence of separately chosen words. If your candidates look like a plain dictionary with no reordering, the grammar model or the tone dictionary did not load, and the first thing to check is whether your Rime build can run Lua at all. If it cannot, switch to Pure.

The custom_phrase.txt and custom_phrase.dict.yaml files at the repository root are the intended place for your own entries. The README does not document their format, so read the existing file contents before editing.

Where Rime Wanxiang is the wrong tool

The clearest limitation is the Lua dependency. Three of the four editions ship Lua extensions, and the README positions Pure as the answer for environments that cannot run Lua. That means on an old Rime build you are choosing between the full experience and a reduced one, and there is no middle path documented.

The dictionary is also a single upstream artefact. The README states that the Wanxiang dictionary has entered a manual maintenance phase, which is a candid admission that the automated optimisation pass is over. Correction of a wrong frequency or a missing term now depends on the maintainers acting on feedback, and the README points to an external feedback form for that purpose. If you need a word added today, you are editing custom_phrase.txt yourself, not waiting for upstream.

Licensing is the third boundary. The project is CC BY 4.0. That is a content licence, not a software licence, and it is applied to a repository that contains Lua code, schema files and dictionary data. The README does not break the licence down by component. If you plan to redistribute the dictionary inside your own scheme, read the licence text and decide for yourself whether attribution requirements fit your distribution; this article is not legal advice.

Finally, if you already maintain a curated dictionary that reflects your own vocabulary, a large general-purpose word list can work against you. Frequency data that suits general writing will sometimes outrank the terms you actually use, and you will spend time fighting the ordering.

How it compares with Mintimate's oh-my-rime

The README's ecosystem section names Mintimate's oh-my-rime, and the relationship is worth understanding before you pick one. Oh-my-rime is described as a comprehensive scheme that uses the Wanxiang dictionary, and its modified Earth Pinyin is said to inherit the Wanxiang tone encoding.

The difference is one of scope. Oh-my-rime is a full scheme in its own right that consumes Wanxiang data as one input among others. Rime Wanxiang is the data and the scheme together, with its own four editions and its own Lua layer. Choosing oh-my-rime means accepting its layout decisions and getting Wanxiang's dictionary as a component. Choosing Wanxiang means the dictionary, the grammar model and the schema come from one maintainer and are versioned together, which makes upgrades more predictable but leaves you with fewer places to swap parts out.

The same pattern holds for the other ecosystem entries. Yuanming Wanxiang, Li's triple pinyin and Wanxiang Hu all build on the Wanxiang ecosystem rather than competing with it, each targeting a layout the main editions do not cover. If your layout appears in that list, start there rather than with Base.

Versioning, upgrades and what a dictionary update costs you

The releases show a numbered line, with v18.0.9 published on 2026-09-21, and a separate dict-nightly release described as a rolling full preview, also published on 2026-09-21. The last push to the repository was on 2026-09-21, so the project is current as of this writing.

The two release channels serve different users. The numbered releases are the stable path. The nightly dictionary is a rolling preview of the full data set, and the README does not describe a rollback procedure for it. If you install the nightly and a term you rely on is reordered or removed, the documented path back is to reinstall a numbered release, not to revert a single file.

Upgrades are a full-directory replacement in most Rime setups, which means your custom_phrase.txt and any local edits sit in the same directory as the files being replaced. The README does not document a merge strategy. Keep your custom entries in a file you can restore, and check the CHANGELOG.md at the repository root before replacing anything, since that is where the project records what changed between numbered versions.

The licence question resurfaces here. CC BY 4.0 governs the repository, and if you redistribute a modified dictionary you are doing so under that licence, with the attribution it requires. Verify the terms on the components you actually ship.

Editorial conclusion

Rime Wanxiang suits Rime users who want stronger sentence-level candidates without assembling a dictionary themselves, and who can run Lua. It is the wrong choice for anyone on a Rime build without Lua, or for users who want a small hand-curated dictionary they can audit line by line; Pure is the fallback for the first case, not the second. Before installing, verify that your Rime front end supports the Lua modules your chosen edition depends on, check which edition matches your input style, and read the licence terms on the dictionary data if you plan to redistribute it.

Frequently asked questions

Official sources

1. amzxyz/rime-wanxiang on GitHub
2. License: CC-BY-4.0
3. Project website
4. README
5. Releases