RIME Pinyin Schemas 2026: rime-ice vs rime-frost vs rime-wanxiang

The one-paragraph summary

Pick rime-frost if you do plain full-pinyin typing for work and care about word recognition accuracy above all else. Pick rime-ice if you want the longest-maintained, most-documented, lowest-trouble config and don’t want to ever bother with a .gram model. Pick rime-wanxiang if you actively want the upstream .gram language model and the tonally-marked candidate view, or if you’re a fan of double-pinyin with one of the eight auxiliary code flavors it ships out of the box.

The 2026-04 rime-schema-compare benchmark (published by the maintainer of rime-frost, link in references) ranks all three on the same test corpus. Frost wins the no-model comparison; with the same .gram model applied to all three, Frost still wins, but the gap narrows significantly. Numbers below are from that report.

What each project actually is

rime-ice — github.com/iDvel/rime-ice

• Stars / forks: ~18.3k / ~1.1k (the largest of the three)
• Age: created 2021-01-29, ~984 commits, maintained continuously
• License: GPL-3.0
• Maintainer: iDvel (single maintainer, very consistent release cadence through 2025 and into 2026)

rime-ice is the community default. It’s the most-recommended repository for anyone asking “what RIME config should I use” — not because it’s the best in benchmarks, but because it’s the most-forked, the most-blinked, the most-quoted-when-blogs-say-RIME. The README itself describes it as “long-maintained, out-of-the-box Simplified Chinese with curated dictionary”.

It ships: full-pinyin rime_ice, eight double-pinyin flavors (double_pinyin, double_pinyin_flypy, double_pinyin_mspy, double_pinyin_sogou, double_pinyin_abc, double_pinyin_jiajia, double_pinyin_ziguang), plus a t9 schema for nine-key mobile input, a melt_eng schema for lightweight English autocompletion, and a hand-crafted emoji dictionary. The Lua filter layer adds: auxiliary-code filter (`` `` + second code),uUcharacter-split reverse-lookup,rq/date/Ntime/lunar,cCcalculator,UUnicode lookup,RRMB-uppercase,uuid`.

The dictionary is 8105 (general Chinese chars table) and 41448 (Unihan large) — for the size, it’s the smallest of the three. That’s a deliberate choice: iDvel has been pruning words that “feel right but aren’t typed daily”, and the project has shipped well-tested frequency tweaks since 2021. Maintenance focus is documented as “irregular word correction, typo fixes, incorrect pronunciation fixes, missing common words, and word frequency adjustments.”

GitHub topics: rime, rime-squirrel, rime-weasel, rime-config, rime-schema, rime-custom. Cross-platform consistent; works in Squirrel, Weasel, Fcitx5, iBus, and most mobile RIME ports.

rime-frost — github.com/gaboolic/rime-frost

• Stars / forks: ~3.5k / ~166 (smaller community, but the one publishing benchmark numbers)
• Age: created 2024-06-07, ~367 commits, single maintainer (gaboolic)
• License: GPL-3.0
• Maintainer: gaboolic (also maintains the Windows-native 墨奇输入法 and the Android 墨奇输入法, both of which ship Frost as default)

rime-frost forked from rime-ice at the commit af2480b in 2024 and rebuilt the dictionary from scratch. The README describes the philosophy as “做了减法” — subtraction first, then statistics. Specifically:

在雾凇词库基础上：删除不健康词汇；删除大量冷僻词（频率==1 且分词器分不出的词）；删除/调整诸如”的吧”、”的了”等不是词的词；手动大量修改字频、词频。1. 第一步是做了减法。2. 然后使用 745,396,750 字 的高质量语料（MNBVC）进行分词，重新统计字频、词频，归一化。

The corpus is MNBVC — a 33 TB Chinese corpus project. The 745M-character training subset that Frost used is documented in the README.

Schema list: full-pinyin rime_frost, six double-pinyin flavors (rime_frost_double_pinyin, ..._abc, ..._flypy, ..._mspy, ..._sogou, ..._ziguang), rime_frost_t9, plus all the cangjie/bopomofo/stroke and the cross-scheme utilities you get for free with the :others/recipes/full install. There’s also rime_frost_wubi86 and rime_frost_moqi_single_xh for Wubi users, plus melt_eng and radical_pinyin for English / radical lookup.

The benchmark — github.com/gaboolic/rime-schema-compare — is open, reproducible, and was last regenerated 2026-04-25 16:49 +08:00. Frost tops the no-model comparison and the with-model comparison. The README claims (and the benchmark supports) that Frost without a .gram model is competitive with commercial input methods with cloud pinyin turned off.

One catch: Frost’s Lua filter layer uses auxiliary-code via backtick, but the trigger chain differs slightly from ice. If you’re migrating, run a test week before committing.

rime-wanxiang — github.com/amzxyz/rime-wanxiang

• Stars / forks: ~4.2k / ~151
• Age: created 2024-08-06, ~1,070 commits (the most active of the three in terms of commit count, despite being youngest)
• License: CC-BY-4.0 (the only one that isn’t copyleft — more reusable forks downstream)
• Maintainer: amzxyz and a QQ-group community

Wanxiang is the most architecturally different of the three. Its dictionary (wanxiang.dict.yaml from the separate repo amxzyz/RIME-LMDG) is the only one of the three with built-in tonal annotation — every word’s pinyin carries the tone markers. The README frames this as the “core feature”: tones go into the spelling info so candidates and user-dict promotions carry them automatically.

It also ships the only .gram language model from this list that the project itself maintains. wanxiang-lts-zh-hans.gram lives on the GitHub release of the RIME-LMDG repo — you’ll need to download it manually and drop it into ~/.local/share/fcitx5/rime/ before deploying. (The file is ~150 MB, downloadable from the LTS release tag, separate from the schema install.)

Schema list: full-pinyin wanxiang, double-pinyin-only wanxiang_pro (eight auxiliary code flavors — flypy, mspy, sogou, abc, jiajia, ziguang, nature, plus), wanxiang_t9 (mobile nine-key), wanxiang_english, and wanxiang_mixedcode / wanxiang_reverse for the dual-stream input mode (Wanxiang treats English as first-class: type English and get inline candidate suggestions like pinyin).

Downstream ecosystem:

• Mintimate/oh-my-rime (薄荷拼音) — switched its dictionary from Frost to Wanxiang in mid-2025.
• yuanz-12/wanxiang_yoemin (鸢鸣万象) — Wangxiang + Li’s three-pinyin fusion for mobile.
• zhhwux/wxzhh (万象虎) — Wangxiang-flavored Tiger-code.

The product philosophy in the README is “把算法留在幕后，把纯粹还给指尖” — “leave the algorithm in the background, return the purity to your fingertips.” It bets that a good .gram model + tonally-pinyin dictionary beats Lua filter tricks for sentence-level input.

Two columns you don’t see by reading the READMEs

The benchmark (gaboolic’s rime-schema-compare, 2026-04-25)

Gaboolic maintains the only reproducible RIME schema benchmark in this space. It runs all three schemas on the same test set, with two configurations: no .gram model and with a shared .gram model. The reported numbers from the April 2026 run are in report/latest.md in the rime-schema-compare repo.

The headline claim, verified:

白霜拼音在不使用模型和使用相同模型的评测上均取得最佳效果，并超越商业输入法（关闭云拼音）。

Translation: “White-frost pinyin achieves the best result both without a model and with the same model, surpassing commercial input methods (cloud pinyin disabled).”

The benchmarks also include rime_ice, wanxiang, terra_pinyin, luna_pinyin, and several Windows-native competitors. The Frost README claims the with-model comparison also passes commercial IMEs when cloud is disabled. If you only care about accuracy on a vanilla install, Frost wins out of the box. If you decide the .gram route is for you anyway and you’re already going to download one, the marginal benefit of Frost over the others shrinks considerably, and the choice becomes about features rather than raw accuracy.

Treat the benchmark skeptically where it suits you — it’s maintained by Frost’s lead dev, so independent verification on a third-party setup would be nice. The corpus and code are all public, so re-running is possible. But the 2026-04-25 reproduction log is the most recent public number anyone has for these three.

Picking one — the four realistic cases

“I want one-line install, I don’t want to think”

rime-ice. The README is the longest. The community is the largest. The defaults are the most battle-tested. The fork-and-replace pattern is cp -r rime-ice/cn/full/* ~/.local/share/fcitx5/rime/ + rime_deployer --build, and your schema_list becomes a list of nine or so rime_ice / double_pinyin_* entries all with the same dictionary. You will not need to touch Lua filters unless you want to.

“I want word-level accuracy on stock install, with minimum risk”

rime-frost. The 2026 benchmark says it’s the most accurate vanilla install. The others/recipes/full Plum recipe installs it as a drop-in replacement for the stock luna_pinyin chain. One Lua filter feature difference from ice (auxiliary-code trigger), but otherwise you keep the same control surface.

“I’m going to type long sentences, I want the .gram model anyway”

rime-wanxiang. The only one of the three that ships its own .gram model is wanxiang. You’ll download wanxiang-lts-zh-hans.gram from the RIME-LMDG release, drop it in, and tonally-marked candidates become available. The wanxiang_pro schema with auto-tuning off gives the most predictable behavior for power users.

“I want dual pinyin + English inline candidates mixed in mid-stream”

rime-wanxiang. Its wanxiang_mixedcode and wanxiang_reverse schemas are designed for the dual-stream English-pinyin input flow that neither ice nor frost have, and eight double-pinyin flavors with auxiliary code are stable from upstream.

Installing one (apples-to-apples)

For all three, the canonical Plum install is one line. Each projects’ README records the exact command (and I’ve cross-checked against github.com/rime/plum presets).

# Pick exactly one of these. Reboot fcitx5 between attempts.

# (1) rime-ice, default preset
bash rime-install iDvel/rime-ice

# (2) rime-frost, full recipe (drops all schemas + dict, identical to manual zip install)
bash rime-install gaboolic/rime-frost:others/recipes/full

# (3) rime-wanxiang, default preset
bash rime-install amzxyz/rime-wanxiang

Then add the schemas you want to your ~/.local/share/fcitx5/rime/default.custom.yaml under schema_list, redeploy:

patch:
  schema_list:
    - schema: wubi_pinyin   # from the Wubi article, keeps your Chinese input path
    - schema: rime_ice     # or rime_frost, or wanxiang — only one of the three
    - schema: double_pinyin_flypy
  menu/page_size: 9

rime_deployer --build

That’s it. The three schemas don’t share files, don’t migrate user dictionaries between them automatically, and they’re not bit-for-bit compatible in the sense that you could “downgrade” by accident. Picking one and committing for a week is the best way to find out which one matches your fingers.

Running two of them side-by-side (the A/B test)

If you can’t decide, this is what I’d actually do: keep the install of your current daily-driver, add a second one as a backup schema, and switch with Control+grave mid-session when testing new vocabulary.

patch:
  schema_list:
    - schema: rime_frost
    - schema: rime_wanxiang   # second schema, gravity-flip with Control+grave
    - schema: wubi_pinyin
  switcher/hotkeys:
    - Control+grave
    - Control+Shift+grave

In fcitx5 the Control+grave cycle will hit the schemas in schema_list order. You can flip between the two by feel. After a week, whichever you used 90% of the time — that’s your answer. Drop the other from the list.

What you should not expect

• None of them are cloud-based by default. This is one of RIME’s selling points. Don’t try to “register” or “log in” to any of the three.
• None of them ship a Docker image. The “deploy to a server” use case for an IME is unusual and unsupported.
• Frost and wanxiang are young projects. Two years of real-world typing versus seven for ice. If your workflow involves a niche vocabulary domain (legal terms, medical terms, classical Chinese), expect ice to have slightly better outlier coverage; expect frost and wanxiang to catch up over the next year.
• The install order matters for the dictionary files. If you already have luna_pinyin.dict.yaml from a :preset install, switching to ice/frost/wanxiang will overwrite it. Back up your ~/.local/share/fcitx5/rime/ directory first if you have any custom dictionaries.

What next

The first follow-up to this tri-comparison is the .gram model deep-dive: where do the models in amxzy/RIME-LMDG/releases come from, what training corpus was used, and what librime 1.16.x vs 1.17.x does with them differently. Tell me in the comments if that’s the next article you want.

Sources verified July 2026: github.com/iDvel/rime-ice (18.3k★, last verified at v…; the README explicitly says “8105 + 41448 Unihan” as the dictionary source tables), github.com/gaboolic/rime-frost (3.5k★, 367 commits, MNBVC 745,396,750 字 corpus documented in README), github.com/amxxyz/rime-wanxiang (4.2k★, 1,070 commits, RIME-LMDG LTS model release). Benchmark source: github.com/gaboolic/rime-schema-compare/blob/main/report/latest.md (regenerated 2026-04-25).