Kria brings the power of large language models directly to your Android device. Chat with AI models running entirely on your phone — no internet connection
required, no data sent to the cloud. Experience the same quality of interaction as cloud-based AI assistants, but with complete control over your data.

🔒 COMPLETELY PRIVATE
Your conversations never leave your device. Kria runs LLM inference locally using llama.cpp, one of the most efficient open-source inference engines
available. This means zero data collection, zero cloud dependency, and total privacy. No account needed, no tracking, no ads — just you and your AI
assistant.

💬 CHATGPT-LIKE EXPERIENCE
• Familiar chat interface with real-time streaming responses — watch answers appear token by token as they are generated
• Full Markdown rendering including headings, bold, italics, lists, blockquotes, and links
• Syntax-highlighted code blocks supporting 20+ programming languages with horizontal scrolling
• Copy any message or code block to clipboard with a single tap
• Regenerate responses when you want a different answer
• Live performance stats displayed after each response: tokens per second, total duration, and token count

🤖 BRING YOUR OWN MODEL
• Built-in Hugging Face integration — search, browse, and download GGUF models directly from the app
• Support for all popular open-source models: LLaMA, Mistral, Phi, Gemma, Qwen, DeepSeek, and many more
• Import any GGUF model file from your device storage using the file browser
• Automatically detects and lists cached models for quick switching

⚙️ FULL CONTROL OVER GENERATION
Fine-tune every aspect of model behavior to get the responses you want:
• Temperature: control creativity vs determinism (0 to 2)
• Top-K: limit candidate tokens for more focused outputs (1 to 100)
• Top-P (Nucleus Sampling): cumulative probability threshold (0 to 1)
• Min-P: minimum probability filter for quality control (0 to 1)
• Repeat Penalty: reduce repetitive text in responses (1 to 2)
• Context window size: set how much conversation the model remembers (512 to 8192 tokens)
• Max response length: control how long responses can be (64 to 2048 tokens)
• CPU thread count: allocate processing power for optimal speed
• Vulkan GPU acceleration: offload computation to your GPU on supported devices
• All parameters are saved per model and persist across sessions

📊 HARDWARE-AWARE OPTIMIZATION
• Automatic detection of your device's CPU cores, architecture, RAM, GPU, and NPU capabilities
• Smart optimization suggestions tailored specifically to your hardware
• One-tap apply for each recommended setting
• Visual display of available resources to help you choose the right model size

💾 CONVERSATION HISTORY
• Unlimited saved conversations with complete message history
• Search across all your conversations by keyword
• Continue any previous conversation exactly where you left off
• Start new conversations at any time
• Switch models mid-conversation without losing context
• Delete conversations you no longer need

⚡ PERFORMANCE OPTIMIZED
• Dual ARM builds: optimized for both v8.2-a and v9-a architectures
• Multi-threaded CPU inference for maximum speed on modern processors
• Real-time performance metrics displayed during and after generation
• Efficient memory management for running large models on mobile hardware

REQUIREMENTS
• Android 14 or later
• ARM64 device (arm64-v8a)
• Sufficient storage for your chosen model (typically 2–8 GB)

Kria is free and open source. Built with llama.cpp and Jetpack Compose.