Skip to content

Releases

Lokey 0.1.3

Choose a tag to compare

@Ninnix Ninnix released this 10 Nov 17:42

Lokey 0.1.3 releases idle model memory and avoids unnecessary background work while keeping the existing models and inference settings.

  • Release prediction, dictation and translation weights after two minutes without use. Each task has its own deadline, so an unused speech model can expire while you keep typing.
  • Cover CPU and supported Hexagon execution, including both Qwen ASR's decoder and audio encoder. Return unused allocator pages to Android after release.
  • Keep models warm through short pauses, protect active inference jobs, and reload on background workers when needed. Typing remains available without a model-loading screen.
  • Start speech-model warming when recording starts, rather than every time the keyboard opens.
  • Stop memory polling when no models are resident, avoid repeated editor resets, and skip prediction/translation refreshes while the keyboard is hidden.

Model weights, quantization, inference/search budgets, prediction delays and CPU/NPU power settings are unchanged. These changes avoid wasted work; battery-life savings have not been measured.

Download lokey-0.1.3-arm64.apk for Android 9 or later on ARM64. It updates previous release-signed versions with the same signing certificate. The AAB is for store distribution; SHA256SUMS and the public signing certificate are included.

Validation includes OnePlus 15 CPU/NPU load, release and reload checks for both text and both dictation models; prediction/cache checks; a real two-minute idle timeout; and paired warm-response measurements. Release APK/AAB signatures, bundled notices and APK alignment are verified. Runtime testing on a 16 KB page-size device remains pending.

CPU execution and the optional Hexagon v81 kernel are included; other DSP versions retain CPU fallback.

Lokey 0.1.2

Choose a tag to compare

@Ninnix Ninnix released this 10 Nov 13:31

Lokey 0.1.2 improves swipe recognition, adaptive letters and typing without downloaded neural models.

  • Expand swipe dictionaries to up to 50,000 words per language across 21 Latin-script languages, using the existing decoder.
  • Improve completion and conservative, opt-in autocorrection without a downloaded text model. Small public context tables cover English, Italian, Spanish, French, German, Portuguese and Turkish; no typing history is learned or stored.
  • Update adaptive letter widths sooner using cached model probabilities, with dictionary/context probabilities available in no-model mode. Letter centers stay fixed, resizing is bounded, and geometry freezes during touches. Adaptive letters remain optional.
  • Restore the wider idle Done key and preserve the familiar Intl dead-key behavior.
  • Prevent primary and secondary key-label overlap across layouts and separate the welcome/privacy actions.

Download lokey-0.1.2-arm64.apk for Android 9 or later on ARM64. It updates previous release-signed versions with the same signing certificate. The AAB is for store distribution; SHA256SUMS and the public signing certificate are included.

The offline engine improves held-out public-corpus top-three next-word coverage from 6.6% to 14.6% across 7,453 probes. This is a controlled corpus result, not measured human typing accuracy. Semantic prediction remains limited; dictation and translation still require their optional models. No-model vocabulary coverage remains limited to the supported Latin-script dictionaries, with native-script input-method conversion unchanged.

Validation includes OnePlus 15 offline typing, installed-model probability checks, swipe, correction, completion, Intl input, privacy and setup; key-label bounds were checked in 5,520 configurations. Human typing benefits from adaptive letters remain unproven. Release APK/AAB signatures, bundled notices and APK alignment are verified. Runtime testing on a 16 KB page-size device remains pending.

CPU execution and the optional Hexagon v81 kernel are included; other DSP versions retain CPU fallback.

Lokey 0.1.1

Choose a tag to compare

@Ninnix Ninnix released this 06 Nov 16:56

Lokey 0.1.1 improves keyboard labels and makes the typing speed test follow your language and layout.

  • Keep only Clipboard and Translate in the keyboard tools menu.
  • Fix Italian keyboard labels, including Accenti, tasti accento and Spazio, and similar translation problems in other languages.
  • Use the opening six lines of Dante's Inferno I in all 39 supported languages, with an additional Traditional Chinese version. Italian uses the original; the other versions are translations or prose adaptations.
  • Keep your selected keyboard layout during speed tests, including native script layouts and right-to-left text.
  • Separate personal records by passage and layout; include the language and layout in shared results.
  • Add README screenshots and exclude local Play Store files from Git.

Download lokey-0.1.1-arm64.apk for Android 9 or later on ARM64. It updates the release-signed 0.1.0 app using the same signing certificate. The AAB is for store distribution. SHA256SUMS and the public signing certificate are included.

Native libraries are unchanged from 0.1.0: CPU execution and the optional Hexagon v81 kernel are included; other DSP versions retain CPU fallback.

Validation covers locale resources, all speed-test languages, the two-item tools menu, scoring and restart, dark and light speed-test screens, and selected Chinese, Persian and Cyrillic layouts on OnePlus 15. APK and AAB signatures and native-library alignment are checked. Native-speaker review of the translations and runtime testing on a 16 KB page-size device remain pending.

Lokey 0.1.0

Choose a tag to compare

@Ninnix Ninnix released this 06 Nov 12:19

First public release of Lokey, a native Android keyboard with local prediction, autocorrection, translation and dictation.

  • One active text model supplies prediction, autocorrection and translation. LFM-only installations can translate; Qwen is optional.
  • One active speech model supplies dictation, including the translation source editor.
  • Complete-word ranking from probabilities, prefix filtering and RAM model-state reuse.
  • Interface translations for 39 languages, selected through Android's language settings.
  • Keyboard layouts, clipboard history, custom colors, saved designs, recorded switch sounds and haptics.
  • Swipe typing, adaptive letters and autocorrection start off; keyboard height starts at 100%.

Install

Download lokey-0.1.0-arm64.apk for Android 9 or later on an ARM64 device. Model files are downloaded separately from Models. Qwen ASR downloads its matching audio encoder with the speech model. Ordinary typing works without model downloads.

This APK includes CPU execution and the optional Hexagon v81 kernel used on the OnePlus 15. Other DSP versions retain CPU fallback. The AAB is provided for store distribution and cannot be installed directly like an APK.

A release-signed APK cannot update a debug-signed installation. Preserve your existing app data before changing installations. Keep using this release certificate for direct APK updates.

Verification and status

The APK and AAB are signed, use the same public upload certificate, and passed signature checks. The archive includes SHA256SUMS and the public certificate; it does not include the private key or signing passwords.

Device validation on the OnePlus 15 covered privacy-policy access, all locale resources, LFM-only translation, model switching and translation-cache identity, normal translation, editor guards, automatic selection, correction, probability/state reuse, the Models screen and setup. Test retries and limits are recorded in the repository's release-preparation report.

This is an initial beta-quality release. Translation quality depends on the model and language; LFM can produce fluent but incorrect translations. Interface translations still need full native-speaker review, and 16 KB page-size runtime testing remains pending.

Privacy and source

Text and microphone audio are processed locally. Optional model downloads expose network request metadata to Hugging Face and its delivery providers. Read the privacy policy and Data Safety disclosure.

Application code is MIT licensed. Models and bundled third-party components retain their own licenses and offline notices.