5.0.0

v5.0.0

New Features

• Translator mode — new "Translator" system prompt with source/target language dropdowns
  (75+ languages now supported !!)
• LaTeX math hints — toggle that instructs the model to use $...$ / $$...$$ notation for
  math expressions
• Markdown rendering — assistant messages now rendered as Markdown (bold, italic, code
  blocks, lists) via the mikepenz multiplatform-markdown-renderer library
• Smart auto-scroll — dragging up while generating pauses auto-scroll; scrolling back to
  the bottom re-enables it

Bug Fixes

• Model unload race condition — unloadModel() is now a suspend function that signals the
  native side to stop, then waits for the generation coroutine to fully exit before freeing
  native memory (5 s safeguard timeout)
• Partial response lost on navigation — switching conversations or starting a new chat now
  saves any in-progress partial response before unloading
• Custom system prompt not persisting — the custom prompt field now correctly restores its
  value when re-opening Settings
• Misc bugs patched

Performance

• Token batching — UI updates batched every 3 tokens, reducing Compose recompositions
  without visible latency
• Stop-sequence scan optimized — only the last 200 characters are scanned per token instead
  of the full response string
• Thread count auto-detected — defaults to availableProcessors() / 2 (clamped 4–8) instead
  of a hardcoded 4

Settings / UX

• Default max tokens raised from 512 → 2048
• Added 3000 / 3500 / 4000 quick-select buttons for max tokens
• isShrinkResources = true enabled for release builds

Build / Dependencies

• AGP 9.1.1, Kotlin 2.3.20, Compose BOM 2026.03.00
• Migrated from kotlinOptions { jvmTarget } to kotlin { compilerOptions { jvmTarget } } DSL
• Dropped kotlin-android plugin (redundant with new compose plugin)
• security-crypto, biometric, Room, coroutines, core-ktx, serialization all bumped to
  stable releases
• CMake ninja path auto-resolved from local.properties / env vars
• Target/compile SDK bumped to 37
• versionCode = 5

OfflineLLM-v4.0.0

Bundled model is now 0.8B parameters.
new security features :
Screenshot Protection
Tapjacking Protection
Accessibility Data Sensitivity
Auto-Lock on Background
Hardware-Backed Keystore for Keys

OfflineLLM-v3.0.0

OfflineLLM now supports Gemma4 models in GGUF format.
Minor bug patches and optimizations applied.

OfflineLLM v2.0.0

OfflineLLM v2.0.0

A fully offline, private AI chat app for Android. All inference runs on-device via llama.cpp. Zero network permissions.

What's New in v2.0.0

• Advanced Sampling Parameters — Full control over Temperature, Top-P, Top-K, Min-P, and Repeat Penalty with slider UI and plain-English explanations
• Context Size Slider — Adjustable from 512 to 16384 tokens
• Text-to-Speech — Read AI responses aloud (speaker icon on assistant messages)
• Chat Search — Search messages within conversations
• Delete Individual Messages — Long-press any message to delete
• Auto-Title Conversations — Chat titles set automatically from your first message
• Theme Selector — System Default / Light / Dark / AMOLED Black
• Accent Colour Picker — 9 colour options
• Thinking Tag Stripping — Hides blocks from reasoning models
• Empty Response Fix — No more blank message bubbles
• Help Screen — Built-in guide for downloading models from HuggingFace
• About Screen — Version info, license, links

Downloads

• OfflineLLM-v2.0.0-release.apk — Install directly on any Android 14+ device
• gemma-3-270m-it-Q4_K_M.gguf — Bundled model, fast on 4GB RAM devices (~300MB)

Install

1. Download the APK and (optionally) a model file
2. Enable "Install unknown apps" in Android settings
3. Install the APK, complete onboarding
4. Import the GGUF model from Settings → Import GGUF Model

Recommended Models

Model	Size	Best For
Gemma 3 270M Q4_K_M	~300 MB	4GB RAM, fast responses
Qwen3.5 0.8B Q4_K_M	~530 MB	4-6GB RAM, good balance
Gemma 3 1B Q4_K_M	~750 MB	6-8GB RAM
Qwen3.5 4B Q4_K_M	~2.5 GB	8GB+ RAM, best quality

OfflineLLM

OfflineLLM v1.0.0 — Initial Release

A fully offline, private AI chat app for Android. All LLM inference runs entirely on-device via llama.cpp. No internet permissions. No cloud. No tracking. - data never leaves the users device.

Features:

• On-device inference with optimized ARM NEON/SVE/i8mm native libraries
• Streaming token-by-token response display
• Import any GGUF model at runtime via file picker
• Multiple conversations with auto-titling and rename
• Chat search and individual message deletion
• Theme selector (System/Light/Dark/AMOLED Black)
• Accent colour picker with 9 colour options
• Configurable system prompts (General, Coder, Creative Writer, Tutor, Custom)
• Temperature, max tokens, and context size controls
• Optional thinking tag stripping for reasoning models
• Encrypted settings via Jetpack Security
• Optional biometric lock
• Chat export/import as JSON
• Built-in help guide for downloading models from HuggingFace
• Zero network permissions — verified in manifest

Recommended models:

• Gemma 3 270M (Q4_K_M) — Fast, works on 4GB RAM devices - included in this APK by default.
• Qwen3.5 0.8B (Q4_K_M) — Good balance for 4-6GB RAM
• Gemma 3 1B (Q4_K_M) — Recommended for 6-8GB RAM
• Qwen3.5 4B (Q4_K_M) — Best quality for 8GB+ RAM

Install: Enable Unknown Sources, then install the APK via file manager or adb install.

<3 JEGLY