v1.4.0

LocalLLM Android v1.4.0 (versionCode 15)

Self-contained on-device LLM app: llama.cpp GGUF + LiteRT LM, encrypted chat storage, model manager with segmented downloads, Ollama/OpenAI-compatible API server, voice mode, MCP tools, first-party glass UI (no Material3), SDengine test scaffold, on-device Whisper STT.

Release notes since v1.3.1:

• Security: exact-host HF token + CORS validation, API auth/limits, fail-closed Keystore crypto, no debug-signed releases
• Stability: serialized inference with fail-fast busy, per-request LiteRT sessions, native-propagated stop, honest API errors/options
• Features: real MCP JSON-RPC client (tools list/call), local Whisper tiny STT with on-device recording, empty chats stay unsaved, SDengine (TEST) runtime selector with loud warnings
• First-party engine scaffold (:engine): GGUF reader, quantizers, BPE, sampler, KV cache, MoE expert pager, reference kernels, transformer, JNI C++ core (libsdengine.so for arm64-v8a + x86_64)
• Verified: 70+ unit tests green, emulator end-to-end (270MB download, load, 98-token云 complete at 6.4 t/s), signed release smoke test

Install: download app-release.apk on your device and install (signed with the project's upload key).
Whisper STT model (~100MB) downloads in-app on first use from the voice screen.