Documentation

Getting Started

First Launch

When you first open AI Troll, you'll see the onboarding screen where you will:

• Learn about the app's capabilities
• Review and accept the terms of use
• Download your first AI model

Downloading Your First Model

• Tap "Find Compatible Models" to browse available models
• Filter by size, architecture, or quantization type
• Select a model suitable for your device (smaller models use less memory)
• Tap "Download" and wait for the download to complete
• The model will be automatically activated once downloaded

Chat Screen

The main chat interface where you interact with your AI assistant.

• Start a Conversation: Simply type your message and tap send
• Clear Chat: Use the menu to clear the conversation history
• Model Info: Tap the model name to see details about the active model
• Toolbar: Access the Translator Mode button (if available) and the overflow menu (⋮) containing New Chat and Model Info

Model Management

Access: Tap the models icon in the navigation bar

Sections:

• My Models: View imported and downloaded models
• Browse: Search HuggingFace for compatible models (rendered in an efficient, paginated list to avoid device freezes)

Actions:

• Set Active: Choose which model to use for chat
• Delete: Remove downloaded models to free up storage
• Import: Add GGUF models from your device storage
• Filter: Filter chips narrow both lists — for example, show only models that support the Translator.

Importing Models

AI Troll supports importing GGUF model files from your device:

• Go to Model Management
• Tap "Import Local Model"
• Select a .gguf file from your device
• The model will be added to your library

Settings

Generation Parameters:

• Temperature (0.1-2.0): Controls creativity. Lower = focused responses, Higher = creative/varied responses
• Top-K (1-100): Limits choices to the K most likely tokens
• Top-P (0.1-1.0): Nucleus sampling - considers tokens until cumulative probability reaches P
• Repetition Penalty (1.0-2.0): Reduces repetitive output
• Max Tokens: Maximum length of AI responses

Download Settings:

• WiFi Only: Restrict downloads to WiFi connections (recommended)
• Allow Metered: Allow downloads on mobile data

Language Settings:

• App Language: Switch between 12 supported languages (English, Arabic, German, Spanish, French, Italian, Japanese, Korean, Portuguese, Russian, Turkish, Simplified Chinese). This localizes the entire interface, including menus, buttons, notifications, and background processes.

Theme Settings

Open Theme Settings from the navigation drawer to change how the app looks. Changes apply immediately.

• Presets: Choose the AITroll, Black & White, or Custom theme. The Custom theme lets you edit 8 core colors; all other UI colors are derived from them automatically.
• Day & Night colors: Every custom color has separate Day and Night values. Night values are derived automatically unless you override them, and low-contrast color pairs are flagged as you edit.
• Display Mode: Follow your device's dark-mode setting, or pin the app to always use Day or Night colors — handy for previewing Night colors while editing.

Translator Mode

Translator Mode provides single-turn, fully offline translation using the active model, on its own dedicated page. It can be opened via the translate button in the chat toolbar.

• Usage: Pick a source language ("From", or Auto-detect) and a target language ("To"), type your text, and tap Translate.
• Swap: Exchanges the source and target languages.
• Clear & Stop: Clear empties both fields; Stop cancels a running translation while keeping any partial text. Output can be selected and copied.

Availability: The translate button appears once the loaded model has been verified as multilingual-capable (approx. 1B+ parameters, instruction-tuned, e.g. Qwen 1.5B+, Gemma 2B+). English-only, base, reasoning, and code models do not qualify. In Model Management, capable models carry a Translator chip and can be filtered. The feature can be disabled entirely under Settings → Features → "Enable Translation Mode".

Privacy & Limits: Translations are not saved, never enter the chat history, and never leave your device. Translation quality varies by model and language pair.

Model Compatibility

Supported Formats

AI Troll supports GGUF (GPT-Generated Unified Format) models. This modern format includes:

• The model weights
• Embedded tokenizer
• Model configuration

Supported Architectures

AI Troll supports popular mobile architectures:

• Gemma & Phi: Google's highly efficient Gemma and Microsoft's Phi open architectures, with verified context length detection.
• Llama 3 & 3.1: Full support for standard instruction-tuned model families.

Quantization Types

Quantized models are optimized for mobile devices:

• Q4_K_M: Smallest size, Good quality, Low memory usage
• Q5_K_M: Small size, Better quality, Medium memory usage
• Q6_K: Medium size, Very Good quality, Medium-High memory usage
• Q8_0: Large size, Excellent quality, High memory usage

Memory Requirements

• Check your device's available RAM before downloading large models
• The app will warn you if a model may be too large for your device
• Close other apps to free up memory for larger models

Tips & Best Practices

• Start Small: Begin with smaller quantized models (1-3GB) to test your device's capabilities
• Manage Storage: Delete unused models to free up space
• WiFi Downloads: Use WiFi for downloading large models to avoid data charges
• Battery: AI processing is CPU-intensive; plug in your device for extended sessions
• Memory: Close background apps for better performance with larger models

Troubleshooting

Model won't load:

• Check if you have enough free RAM
• Try a smaller quantized version of the model
• Restart the app

Slow responses:

• Use a smaller/more quantized model
• Close background apps
• Ensure your device isn't overheating

Download fails:

• Check your internet connection
• Ensure you have enough storage space
• Try downloading a smaller model first