remiai-ship-it/RemiAI-Edge-Gallery-Installer

RemiAI Edge Gallery

🚀 Download, run, and experiment with open-source AI models locally - No cloud, no API keys, just plug and play!

What is RemiAI Edge Gallery?

RemiAI Edge Gallery is a Windows desktop application that lets anyone easily download and run open-source Large Language Models (LLMs) directly on their laptop. Perfect for:

• 🎓 Students learning about AI without complex setup
• 🔬 Researchers experimenting with different models offline
• 💻 Developers testing LLMs locally before deployment

Features

Core Features

• Model Gallery - Browse and download models categorized by RAM requirements
• 8GB RAM Models - Lightweight models under 1B parameters
• 16GB RAM Models - More powerful models up to 7B parameters
• One-Click Download - Direct download from Hugging Face (no account needed)
• Model Switching - Switch between models seamlessly during chat
• Offline Chat - Everything runs locally on your device

Chat Features

• Conversation Memory - Remembers up to 20 messages for context
• Speed Mode - Fast responses without conversation memory (each reply is independent)
• Think Mode - Optional chain-of-thought reasoning with collapsible thinking display
• AI Characters - 7 built-in personalities + custom character with your own system prompt
• Markdown Support - AI responses render with proper formatting

Custom Model Upload

• Upload Zones - Large drag-and-drop upload areas in both Model Gallery and My Models views
• Upload Your Own Models - Import any .gguf model file from your computer
• GGUF Validation - Automatically validates model format before import
• Easy Integration - Custom models appear in your downloaded models list

Storage & Disk Management

• Storage Directory Picker - Choose where to store model weights on first launch
• Change Storage Location - Move all models to a new directory at any time
• Real Disk Space Detection - Accurate free space display using PowerShell
• Low-Space Alerts - Warning before downloading when disk space is low (doesn't block)
• Permanent Deletion - Models are permanently deleted, not sent to recycle bin

UI Features

• Dark/Light Theme - Toggle between themes
• System Info Display - Shows RAM usage, disk space, and storage path
• Download Progress - Real-time download progress with cancel option
• Confetti Animation - Celebration when downloads complete!

Quick Start

Prerequisites

• Windows 10/11
• Node.js 18+ installed
• 8GB RAM minimum (16GB recommended for larger models)

Installation

# Clone the repository
git clone https://huggingface.co/spaces/remiai/edge-gallery
cd edge-gallery

# Install Git LFS (required for engine binaries)
git lfs install
git lfs pull

# Install dependencies
npm install

# Start the application
npm start

First Run

1. Open the app - you'll be asked to choose a storage folder for model weights
2. You'll see the Model Gallery with upload zones at the top
3. Choose a model based on your RAM (8GB or 16GB tab)
4. Click Download and wait for it to complete
5. Click Use Model to start chatting!

Upload Custom Model

1. Go to My Models tab
2. Click Upload Custom Model button
3. Enter a name for your model
4. Click Browse to select your .gguf file
5. Click Import - model is copied and ready to use!

Available Models

8GB RAM (Under 1B Parameters)

Model	Parameters	Size	Category
Llama-3.2-1B-Instruct	1B	810 MB	Instruct
SmolLM2-360M-Instruct	360M	290 MB	Chat
SmolLM2-135M-Instruct	135M	120 MB	Chat
Qwen2.5-0.5B-Instruct	0.5B	397 MB	Instruct
Danube3-500M-Instruct	500M	380 MB	Chat
TinyLlama-1.1B-Chat	1.1B	638 MB	Chat

16GB RAM (Under 7B Parameters)

Model	Parameters	Size	Category
SmolLM2-1.7B-Instruct	1.7B	1.1 GB	Chat
Qwen2.5-1.5B-Instruct	1.5B	1.1 GB	Instruct
Qwen2.5-3B-Instruct	3B	2.0 GB	Instruct
Gemma-2-2B-IT	2B	1.5 GB	Instruct
Phi-3.5-mini-instruct	3.8B	2.4 GB	Instruct
Phi-3-mini-4k-instruct	3.8B	2.4 GB	Instruct
Llama-3.2-3B-Instruct	3B	2.0 GB	Instruct
StableLM-Zephyr-3B	3B	1.8 GB	Chat
Phi-2	2.7B	1.6 GB	Code
DeepSeek-Coder-1.3B	1.3B	990 MB	Code
CodeGemma-2B	2B	1.5 GB	Code
Rocket-3B	3B	1.7 GB	Chat
Danube2-1.8B-Chat	1.8B	1.2 GB	Chat
Yi-1.5-6B-Chat	6B	3.6 GB	Chinese
Qwen2-1.5B-Instruct	1.5B	1.1 GB	Chinese

Model Settings

Customize your chat experience with these settings:

• Context Memory - Total conversation memory window (512-8192 tokens)
• Max Input Tokens - Maximum length of your input messages (100-4096 tokens)
• Max Output Tokens - Maximum response length (100-4096 tokens)
• Temperature - Creativity level (0 = focused, 2 = creative)
• Speed Mode - No conversation memory — each reply is independent (fastest responses)
• Think Mode - Shows the model's step-by-step reasoning in a collapsible block above the answer
• AI Character - Choose personality type:
  • 🤖 Helpful Assistant - General assistance
  • 📚 AI Tutor - Step-by-step explanations with examples
  • 💻 Code Helper - Clean, well-commented programming assistance
  • ✨ Creative Writer - Imaginative, expressive writing help
  • 🔬 Scientist - Evidence-based scientific explanations
  • 🌍 Translator - Accurate language translation
  • 🧘 Philosopher - Deep philosophical discussions
  • ✏️ Custom Character - Write your own system prompt!

File Structure

remiai-edge-gallery/
├── main.js           # Electron main process, engine management, IPC handlers
├── preload.js        # Secure bridge between main and renderer
├── renderer.js       # UI logic, chat functionality, model management
├── models.js         # Model catalog with verified download URLs
├── index.html        # Application UI layout
├── styles.css        # Modern dark/light theme styling
├── engine/           # LLM engine binaries (llama.cpp based)
│   ├── cpu_avx/      # AVX version (broader compatibility)
│   ├── cpu_avx2/     # AVX2 version (faster if supported)
│   └── models/       # Downloaded model files (.gguf)
├── package.json      # Project configuration
├── README.md         # This file
├── document.md       # Student/Developer documentation
└── report.md         # Technical architecture report

How It Works

1. Engine: Uses bujji_engine.exe (llama.cpp based) for CPU inference
2. Format: All models use .gguf format (quantized for efficiency)
3. API: Engine runs a local server on port 5000 with OpenAI-compatible API
4. Download: Models downloaded directly from Hugging Face (no login required)
5. Storage: Models saved in user-chosen directory (configurable, defaults to AppData)
6. Custom Models: Users can import their own .gguf files via upload zones or browse dialog
7. Disk Space: Detected via PowerShell Get-PSDrive with fs.statfsSync fallback
8. Deletion: Models permanently deleted via fs.unlinkSync with verification

Troubleshooting

Model won't load

• Check if the .gguf file downloaded completely
• Ensure you have enough free RAM
• Try a smaller model first
• Wait for "Ready" status (larger models take longer to load)

Slow responses

• Enable Speed Mode in settings
• Disable Think Mode if you don't need the reasoning display
• Use a smaller model (< 1B parameters)
• Close other applications to free RAM
• Reduce Max Tokens in settings

Download fails

• Check your internet connection
• Try downloading again (downloads resume from where they stopped)
• Some models may require Hugging Face authentication

"Object destroyed" error

• Fixed in latest version - update your code
• The app now safely handles window close during operations

Engine error during loading

• Wait 1-2 minutes for large models to load
• The engine takes time to load larger models into memory

Building for Distribution

# Create Windows installer
npm run dist

The installer will be created in the release/ folder.

Tip: Run PowerShell as Administrator if you encounter build permission errors.

License

MIT License - Free for educational and personal use.

---

Made with ❤️ by RemiAI - Making AI education accessible to everyone!