PrivateLM

A production-ready, cross-platform AI chat client built with Flutter. It unifies local on-device LLM inference (Android) with cloud API access, giving users full control over how their models run.

Image generation tested on Moto G71 (Snapdragon), Oneplus 10r (Mediatek), Pixel 6A (Tensor), Poco F1 (Snapdragon), Samsung s23 (Snapdragon) 4 steps fast

Generated on pixel 6 with 20 step

---

What It Does

• Local Inference on Android — Download and run GGUF models directly on your phone using GPU-accelerated inference (Vulkan). No internet required after download.
• Cloud API Fallback — Seamlessly switch to OpenAI, Anthropic, Google Gemini, or Kimi (Moonshot AI) when you need more power or are on unsupported platforms.
• Multimodal Chat — Send text and images in conversations. Vision support works with both local models (Qwen2-VL) and cloud providers.
• Persistent Sessions — All chats and settings are stored locally via Hive. Nothing leaves your device unless you explicitly choose cloud mode.
• Background Services — Firebase Cloud Messaging integration for push updates and background task handling.
• Smart Auto-Configuration — On first launch, the app detects your device's RAM and recommends optimal context size and token limits automatically.
• Task Management (in development) — A task view for structured, multi-step AI workflows is being built. It is not available in the app yet.

---

Technical Architecture

Stack

• Framework: Flutter 3.x (Dart >=3.3.0)
• State Management: GetX
• Local Storage: Hive
• Networking: Dio + package:http
• Background Execution: flutter_background_service + flutter_local_notifications
• Push Notifications: Firebase Core + Firebase Messaging

Inference Pipeline

┌─────────────────────────────────────────────────────────────┐
│                        UI Layer                              │
│   ChatView / ModelView / ServerView / SettingsView          │
└──────────────────────────┬──────────────────────────────────┘
                           │
┌──────────────────────────▼──────────────────────────────────┐
│                    Controllers (GetX)                        │
│   ChatController · TaskController · ModelController         │
│   SettingsController · HomeController                       │
└──────────────────────────┬──────────────────────────────────┘
                           │
┌──────────────────────────▼──────────────────────────────────┐
│                      Services                                │
│  ┌─────────────────┐  ┌─────────────────┐  ┌─────────────┐ │
│  │ InferenceService│  │  CloudService   │  │DownloadSvc  │ │
│  │  (local GGUF)   │  │ (OpenAI/Claude/ │  │ (model dl)  │ │
│  │                 │  │  Gemini/Kimi)   │  │             │ │
│  └─────────────────┘  └─────────────────┘  └─────────────┘ │
│  ┌─────────────────┐  ┌─────────────────┐  ┌─────────────┐ │
│  │  HiveService    │  │ DeviceInfoSvc   │  │ExecutionSvc │ │
│  │  ( persistence) │  │  (RAM/GPU tier) │  │ (bg tasks)  │ │
│  └─────────────────┘  └─────────────────┘  └─────────────┘ │
└─────────────────────────────────────────────────────────────┘

Local Inference (Android)

The app uses llama_flutter_android, a custom Flutter plugin wrapping llama.cpp for ARM64 devices. At runtime it:

1. Detects GPU capabilities via Vulkan to determine offload layers.
2. Selects thread count based on device tier (ultra / high / mid / low).
3. Loads the GGUF model with progress streaming.
4. Generates tokens via generateChat() with native chat-template support (ChatML, Llama-3, Gemma, Phi).
5. Falls back to manual prompt construction if native templates fail.

Idle detection (5s) and hard timeouts (180s) keep the UX responsive even on underpowered hardware.

Cloud Inference

CloudService normalizes four different API shapes into a single interface:

• OpenAI — standard /v1/chat/completions
• Anthropic — Messages API with separate system param
• Google Gemini — generateContent with inline image base64
• Kimi — OpenAI-compatible endpoint from Moonshot AI

API keys are stored in Hive and never transmitted anywhere except to the provider's endpoint.

Cross-Platform Abstraction

Local inference is conditionally compiled:

• Android → inference_android.dart (full llama.cpp engine)
• Web → inference_stub.dart (cloud-only, local coming soon)
• iOS → inference_android.dart (full llama.cpp engine via Metal GPU)

The InferenceService exposes supportsLocalInference so the UI can hide local-model UI on unsupported platforms.

---

Supported Platforms

Platform	Local Inference	Cloud APIs	Notes
Android	✅ Yes	✅ Yes	CPU offload via NEON; minSdk 28
iOS	✅ Yes	✅ Yes	Metal GPU acceleration
Web	❌ No	✅ Yes	Cloud-only (local coming soon)

iOS / iPad (Installation via AltServer)

The iOS/iPad release is distributed as a sideloadable IPA file. Due to the RAM requirements for local models, iPad is the recommended iOS target, but it can be installed on iPhones.

Here is how you can install the app using AltStore and AltServer:

Prerequisites:

• Ensure AltServer is installed and running on your Mac or Windows PC.
• Your iOS device must be connected to the same Wi-Fi network as your computer, or connected directly via a USB cable.
• You must have AltStore successfully installed on your iOS device.

여기까지가 README 앞부분입니다. 나머지는 원문에서 읽을 수 있습니다. README 전문 보기 ↗