Your AI. Your Device.
MobiMind is a privacy-focused, multi-model mobile AI platform designed for completely local LLM inference, native tool execution, and verifiable automated task completion on Android.
MobiMind separates natural human dialogue from tool calling using a high-efficiency Dual-Engine Architecture:
- Conversational Brain: The user's chosen LLM (Gemma 3 1B, Qwen 2.5 Coder 1.5B, Llama 3.2 1B, etc.) generates creative and conversational answers.
- Action Engine: Google FunctionGemma 270M runs locally to translate actionable intents into structured tool execution declarations.
- Verification Loop: Verifies tool outputs, gathers evidence, and presents confirmed results back to the user.
User Input │ ▼ Selected LLM │ ▼ Task Classifier ├── CHAT │ │ │ ▼ │ Selected LLM (Direct Response) │ └── ACTION │ ▼ FunctionGemma 270M (Action Engine) │ ▼ Tool Executor (Native Android APIs) │ ▼ Verification Engine │ ▼ Evidence Collector │ ▼ Selected LLM (Synthesis) │ ▼ Final Response
- 🔒 100% Private & Offline: Zero telemetry. All inference, memory storage, and tool calls execute entirely on-device.
- ⚡ Multi-Model Support: Native support for cutting-edge GGUF v3 models optimized with 2-bit (Q2_K) and 4-bit (Q4_K_M) quantization.
- 🛠️ Function Calling on Edge: Dedicated action engine powered by FunctionGemma 270M.
- 📱 RGB Neon Liquid Glass UI/UX: Translucent glass cards, glowing gradient borders, fluid animations, and a modern 5-tab navigation.
Model metadata and download specifications are maintained in config/models.json:
| Model | Size | Quant | Format | Best For |
|---|---|---|---|---|
| Google FunctionGemma 270M | ~241 MB | Q4_K_M | GGUF | Tool calling & Action Engine |
| Google Gemma 3 270M | ~241 MB | Q4_K_M | GGUF | Ultra-compact general dialogue |
| Google Gemma 3 1B | ~658 MB | Q2_K | GGUF | Edge NLP & Reasoning |
| Qwen 2.5 Coder 1.5B | ~645 MB | Q2_K | GGUF | Code synthesis & Troubleshooting |
| Meta Llama 3.2 1B | ~554 MB | Q2_K | GGUF | Fast conversational text |
| IBM Granite 4.0 1B | ~562 MB | Q2_K | GGUF | Enterprise RAG & Context |
| SmolLM2 1.7B | ~643 MB | Q2_K | GGUF | Reasoning & general chat |
| Falcon3 1B | ~693 MB | Q2_K | GGUF | Scientific & analytical QA |
- Node.js v20+ / v22 LTS & Yarn
- JDK 21 (JAVA_HOME pointing to OpenJDK 21)
- Android SDK (API 29+) & NDK 27+
�ash yarn install cd android ./gradlew assembleProdDebug
The resulting APK will be located at:
ext android/app/build/outputs/apk/prod/debug/app-prod-debug.apk
�ash adb install -r android/app/build/outputs/apk/prod/debug/app-prod-debug.apk
MobiMind is continuously tested in Waydroid (Android 13 x86_64) under WSL2: `�ash
adb push ./models/q2-test/google_gemma-3-1b-it-Q2_K.gguf /data/user/0/com.pocketpalai/files/models/local/
waydroid session start waydroid app launch com.pocketpalai `
- llama.rn / llama.cpp local inference integration
- 6-model 2-bit GGUF evaluation and benchmark suite
- FunctionGemma 270M tool calling engine
- Centralized model catalog (config/models.json)
- RGB Neon Liquid Glass design system & UI components
- Cactus / Needle 2 C++ engine support for .cact models
- Hardware-accelerated NPU/DSP delegates (QNN / MediaTek NeuroPilot)
This project originated from and builds upon the open-source PocketPal AI project by Asghar Ghorbani and contributors, licensed under the MIT License.
MobiMind introduces the Dual-Engine Action/Chat architecture, FunctionGemma agent verification pipelines, multi-model Q2_K catalog, and RGB Neon Liquid Glass design system.
See the LICENSE file for complete license terms.
2026 - All Rights Reserved