Your Private AI, Completely Offline.
Run Llama 3, DeepSeek-R1, Mistral, Phi-3 & Gemma on your phone no cloud, no tracking, no compromise.

PocketLLM Lite is an ultra-premium, privacy-first offline AI chat application for Android, iOS, and Desktop. By seamlessly bridging local on-device inference with Ollama via Termux (Android) or native macOS/Linux pipelines, it enables running state-of-the-art Local Large Language Models (LLMs) like DeepSeek-R1, Llama 3.2, and Phi-3 completely offline safeguarding your digital sovereignty with zero external data telemetry.

Designed with an exquisite, fluid Material 3 Expressive Design System, PocketLLM Lite delivers lightning-fast offline completions, advanced Retrieval-Augmented Generation (RAG) document ingestion, local TTS/STT, a native agentic tool execution system, and a robust open-standard Agent Skills runtime!

PocketLLM Lite now includes a robust Agent Skills System built on standard SKILL.md open formats. You can expand your offline assistant's capabilities dynamically:

• One-Tap GitHub Installer: Install complex agent skills directly from any GitHub or raw SKILL.md URL. The engine handles redirects to raw content and parses YAML frontmatter and markdown instructions instantly.
• Manual CRUD Creator: Author and design customized skills offline using premium Material 3 sheets and input validation.
• Smart Autocomplete Suggester: Typing / inside the chat bar triggers a horizontal recommended skills chip panel.
• In-Input Highlights & Cursor Navigation: Skill trigger words like /webdesign are highlighted in bold blue inside your text field. Tapping a highlighted token redirects you to its detailed instructions.
• Dynamic Context Injection: The inference pipeline automatically scans your query, reads matching active skill markdown instructions, and injects them into system prompts before generating a reply.

• Complete Offline Completion: 100% on-device inference with zero cloud dependencies. No chat logs, API credentials, or behavioral metrics ever leave your phone.
• DeepSeek-R1 Chain-of-Thought Accoridon: Native streaming support for <think> reasoning tags, rendered in a beautifully animated Material 3 collapsible accordion.
• Dynamic AI Personas: Save system instructions with distinct emoji avatars, temperature overrides, and default LLM bindings. Includes a scrollable persona switcher HUD.
• Knowledge Base RAG: Process and query offline vector store databases (PDF, TXT, CSV) to automatically augment prompts with secure local contexts.
• Vision & Multimodal Support: Attach camera snapshots or gallery images directly to multimodal LLMs (e.g. Llama 3.2 Vision).

Enable agentic capabilities to let local models execute offline code tools:

• Offline Calculator: Solves complex and basic mathematical equations.
• System Diagnostics: Queries native OS platforms, dates, times, and zones.
• Offline Knowledge Search: Provides simulated general reference knowledge.

• Dictation (STT): Dictate your prompts offline using an animated bottom microphone HUD.
• Speak Back (TTS): Single-tap speak chips to read any AI completion aloud using offline native speech engines.

To run the offline local inference engines directly on your Android device:

1. Install Termux: Download from F-Droid (the Google Play version is obsolete).
2. Install & Setup Ollama:

   pkg update && pkg upgrade
   pkg install ollama

3. Launch Ollama Server:

   ollama serve

4. Download Local Models (open a new Termux session):

   # Pull DeepSeek R1 1.5B Reasoning Model
   ollama pull deepseek-r1:1.5b
   
   # Pull Llama 3.2 3B Text Model
   ollama pull llama3.2

5. Connect PocketLLM: Launch the app, navigate to Settings, configure your endpoint URL (usually http://127.0.0.1:11434), and tap Connect.

flutter pub get

dart run build_runner build --delete-conflicting-outputs

flutter run

# Windows
build_release.bat

# Linux/macOS
chmod +x build_release.sh && ./build_release.sh

# Manual Execution:
flutter build apk --release --obfuscate --split-debug-info=./debug_symbols

• Dart Obfuscation & R8/ProGuard: Production builds are heavily obfuscated to strip debug symbol indices, rename classes, and compress asset footprints.
• Secure Local Box Sandbox: Chat database boxes are stored in native application sandboxes, protected by OS-level permissions.
• Sanitized Telemetry: Legacy advertising monetization libraries, Google AdMob SDK hooks, and tokens usage trackers are completely removed, establishing a clean offline client.
• Cleartext Permissions: Only network-loopback permissions (for Ollama communication at localhost:11434), camera access, and sandbox reads are requested.

• Logic Framework: Flutter & Dart Stable Engine
• State Management: Riverpod 3 (StateNotifiers & async NotifierProviders)
• NoSQL Local Storage: Hive CE (Community Edition, fast binary serialization)
• Nav Systems: GoRouter with nested ShellRoutes and animation transitions
• Design Foundations: Material Design 3 (Seed HSL colors, dynamic color schemes)
• Streaming: Direct custom HTTP stream parser supporting Chunked Transfer encodings

lib/
├── core/               # Global configurations, router, and themes
│   ├── constants/      # App-wide constants, system presets, and prompts
│   ├── theme/          # Dynamic Material 3 styling engine
│   └── widgets/        # Shared core widgets (M3AppBar, M3Avatar, M3EmptyState)
├── features/           # Feature-driven domain modules
│   ├── chat/           # Prompts inputs, bubbles, and autocomplete logic
│   │   ├── domain/     # Data models (ChatMessage, ChatSession, Skill, Persona)
│   │   └── presentation/ # Chat UI, management screens, and state providers
│   ├── media/          # Global media gallery grids
│   ├── profile/        # Personalized user profile configurations
│   ├── settings/       # Appearance sliders, benchmarker, and updates HUD
│   ├── tags/           # Tag organization screens
│   └── splash/         # Native launch and loading handlers
└── services/           # Underlying storage, Ollama, and update layers

PocketLLM Lite is licensed under the permissive MIT License—enabling absolute freedom to copy, inspect, modify, fork, or redistribute.

If you encounter performance issues with local weights execution, have questions regarding setup configurations, or would like to contribute:

1. File a bug report in the development workspace.
2. Include your specific device specifications (RAM/CPU/GPU) and active Termux packages.
3. Reach out to the core engineering team.

PocketLLM Lite is crafted with Flutter. Offline AI for everyone.