Repository navigation
Lokey 0.1.3
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.