mrkeyoor.com_
Sun 04 Oct 07:02 UTC
Dev Toolsevaluationupdated 04 Oct 2026

qingjian review

QingJian is a Chinese-first input method that places an English, Japanese, or Spanish translation beside each Chinese candidate as you type. Its README and user and developer documentation are in Chinese, and we found no English documentation. It runs its pinyin conversion, dictionaries, local sentence model, learning, and usage statistics on the device.

Verdict

Our QingJian build failed in 62 seconds and its tests failed in 35 seconds because the Debian workspace tried to compile Apple-only objc2, so Linux source validation is not clean at commit c08ae57. The product idea is unusually focused: learn one translated word while doing normal Chinese input, with local processing as the default. Try the packaged app only if Chinese documentation, unsigned installers, and beta-grade platform gaps are acceptable.

We ran it

Lab card: what happened when we ran qingjianScreenshot of qingjian (qingjian.app)
Install✓ · 37s442 packages
Build✗ · 62s
Tests✗ · 35sran, no count parsed
Repo965 files~85,047 lines of source · 64.1 MB · 3 CI workflows

Answers from our run

Does qingjian build from source?

Dependencies installed in 37 seconds (442 packages), and the build failed. We cloned commit c08ae57 into a clean Debian container with 3 CPUs and no project-specific setup.

Do qingjian's tests pass?

The test command failed in our container, and its output did not report a pass or fail count.

Who should not use qingjian?

English-only users or contributors: the project states that documentation and comments are written in Chinese, and we found no English guide.

What are the alternatives to qingjian?

librime, Fcitx 5, Fcitx 5 Rime. Our QingJian build failed in 62 seconds and its tests failed in 35 seconds because the Debian workspace tried to compile Apple-only objc2, so Linux source validation is not clean at commit c08ae57.

Setup2/5Debian build hits Apple-only objc2; installers are unsigned
Docs3/5Detailed Chinese guides, but no English documentation found
Community4/53,923 stars with 136 issues and 36 PRs open
Maturity2/5v0.1.4 is beta software with visible platform gaps

Who it’s for

Chinese typists who want vocabulary prompts woven into ordinary writing.
macOS 13+ or Windows 10/11 users willing to test a young input method.
Linux users comfortable installing an Fcitx5 component and starting its service manually.
Rust contributors who can work from Chinese technical documentation and test native platform shells.

Who it’s NOT for

English-only users or contributors: the project states that documentation and comments are written in Chinese, and we found no English guide.
Security teams that forbid unsigned desktop software: v0.1.4 lacks Apple developer signing and Windows code signing.
Linux users expecting feature parity: the README says the Fcitx5 build has no cloud suggestions or settings interface and needs a manually started service.
Anyone unwilling to send typed context to an AI vendor: optional cloud suggestions send the current input and nearby text directly to the provider you configure.
Debian CI jobs expecting a platform-neutral workspace build: our build and tests stopped on the Apple-only objc2 crate.

Setup reality

Our Debian sandbox installed commit c08ae57 in 37 seconds and pulled 442 packages. The build failed with exit 101 after 62 seconds, and the test command failed with exit 101 after 35 seconds. Both logs identify the same blocker: objc2 only compiles for Apple platforms.

The packaged route depends on the desktop. The README supports macOS 13+, Windows 10/11, and Linux through Fcitx5. Release v0.1.4 says the macOS and Windows installers are unsigned; Linux has an Ubuntu 26.04 x86_64 package, while other distributions need a source install.

No account is required for local typing. Cloud suggestions are off by default and need your own AI-provider configuration. Enabling them sends the current input and nearby text to that provider, so the privacy decision changes with one setting.

Three learning languages sit beside normal Chinese candidates

QingJian adds one small thing to a familiar input-method window: a translation beside each Chinese candidate. The user can choose English, Japanese, or Spanish, or turn translations off. Pinyin, full-sentence input, abbreviated pinyin, spelling correction, Shuangpin, Zhuyin, and Wubi are all described in the README. The translation remains secondary to the text you are trying to enter.

Typing and language study share the same moment. QingJian also records today's input volume, the last 7 days, lifetime use, and encountered vocabulary. Those statistics stay on the device. The project says pinyin conversion, dictionaries, the local sentence model, learned habits, and input statistics all run locally, with no account required.

Chinese documentation is a real adoption boundary

The repository's README, user guides, development notes, and code comments are written in Chinese. Its contributing guide says this is deliberate, while code identifiers and error strings remain English. We found no English README or English documentation tree. An English-speaking user can translate pages, but installation, privacy details, troubleshooting, and architecture decisions are written for a Chinese reader.

That choice matches the audience, yet it narrows who can safely contribute to an input method. Keyboard software touches password fields, accessibility APIs, and operating-system text services. QingJian's Chinese technical notes are unusually detailed, including platform boundaries and performance records. The language barrier, rather than missing effort, is what blocks an English-only contributor.

What happened when we ran it

Our run installed 442 Rust packages in 37 seconds at commit c08ae57. The fresh Debian container had 3 CPUs, 12 GB of RAM, no secrets, and no elevated privileges. Its checkout contained 965 files, about 85,047 lines of source, and occupied 64.1 MB. Our test method ran the workspace on Linux rather than using a macOS runner.

The build failed with exit 101 after 62 seconds. Cargo reached objc2 v0.6.4, whose compile-time message says the crate only works on Apple platforms and asks for an Apple target, unless the caller is intentionally using GNUStep with its feature. The log then reports that objc2 could not compile. It does not show an application-level Rust error or a missing Debian system package.

The test command failed the same way after 35 seconds. It also reached objc2 v0.6.4 and stopped on the Apple-platform check before a test summary appeared. QingJian has 3 CI workflow files and no top-level tests directory, but those repository signals do not turn our failed Linux run into a pass. The default workspace path we measured is not platform-neutral on Debian.

macOS and Windows packages still carry beta warnings

Release v0.1.4 was published on September 24, 2026, and labels the macOS and Windows builds as testing-stage software. The macOS package has no Apple developer signature, so Gatekeeper requires approval under Privacy and Security. The Windows installer has no code signature and can trigger SmartScreen. Those warnings deserve attention for software that reads every composing keystroke.

The issue queue shows why the beta label matters. Issue 347 reports blank Chinese candidates on macOS 27 after a font-outline change. Issue 356 covers translation-selection shortcuts failing in Ghostty and cmux. A Windows issue describes the candidate window separating from the text field on a dual-monitor game setup. These failures sit in the system surfaces an input method must get right.

Linux has Fcitx5 support without the full settings experience

QingJian's first Linux build works through Fcitx5 and uses the system candidate panel. Release v0.1.4 provides a prebuilt x86_64 package made on Ubuntu 26.04. Other distributions are directed to build from source. The background service needs manual startup, and the README says Linux lacks cloud suggestions and a settings interface. Local candidates, learning, translations, and sentence-model reranking are present.

That is enough for an experienced Fcitx5 user to experiment. Configuration-file edits and service restarts replace parts of the graphical workflow. Our Debian compilation result adds another caution: the root workspace build attempted an Apple-only dependency. Use the documented Linux package or Linux-specific instructions rather than assuming a plain Cargo workspace command will work.

Cloud suggestions change the privacy model when enabled

Local operation needs no account and uploads no typing statistics, according to the README. Update checks contact the official site for a version list and can be disabled. Input logs are stored separately from statistics and can be turned off or cleared. Those choices give a user a reasonable local baseline for software that sits in the typing path.

Cloud suggestions are disabled by default. Once enabled, QingJian sends the current input and nearby text directly to the AI provider whose credentials you supply. Requests do not pass through QingJian's server, but the provider still receives context from what you are writing. Version v0.1.4 also changed logging so loaded API keys are replaced with *** and malformed environment lines no longer expose a key in logs.

Active October work does not erase the platform risk

GitHub showed 3,923 stars, 136 open issues, and 36 open pull requests on October 4, 2026. The last repository push was September 29, five days earlier, and recent work covers all 3 supported desktop systems. That is a busy project created on September 7. It is also too young for popularity to substitute for long-term input-method reliability.

QingJian is worth trying if the translation-in-the-candidate-window idea matches how you learn and you can read its Chinese support material. Keep cloud suggestions off until you have reviewed the provider's data handling, and test password fields plus your daily applications before making it the default. The 442-package install was easy in our sandbox. The platform-coupled build and unsigned releases make a fallback input method necessary.

Alternatives

ProjectWhat it isPick it when
librimeThe core engine behind the Rime family of configurable input methods.pick this instead when mature schema customization and a broad front-end ecosystem matter more than built-in translation prompts.
Fcitx 5A Linux input-method framework with mature desktop integration.pick this instead when Linux integration and framework stability matter more than QingJian's language-learning layer.
Fcitx 5 RimeA bridge that runs Rime input schemes inside Fcitx 5.pick this instead when you want Rime's schemas on Linux without QingJian's early Fcitx5 limitations.

What people are saying

  1. [velocity-scout] qingjian-team/qingjian

Sources

  1. QingJian repository
  2. QingJian README
  3. QingJian v0.1.4 release
  4. QingJian documentation policy
  5. macOS 27 candidate-window issue 347
  6. Terminal shortcut issue 356

More dev tools reviews

github-stars-history · BrokenPipe · FGOAC-scooby · plexo · window-sweaters · airlift · the whole board →