# Lapis: Complete Technical & Operational Documentation for LLMs

## Overview
Lapis is an offline, privacy-first artificial intelligence assistant engineered natively for Apple Silicon devices (iOS and iPadOS). It enables users to run modern LLMs, reasoning models, and vision-language models (VLMs) directly on their hardware without requiring an internet connection or transmitting data to third-party cloud servers.

## Technical Architecture

### 1. Inference Engine
- **Framework**: Apple MLX Framework with Metal Performance Shaders (MPS).
- **Compute Layer**: Direct execution on the Metal 3 GPU and Apple Neural Engine (ANE).
- **Unified Memory Advantage**: Eliminates PCIe data transfer bottlenecks between CPU and GPU RAM. Weight matrices and KV Cache reside directly in shared LPDDR5/LPDDR5X unified memory.
- **Quantization Support**: 4-bit and 8-bit quantization formats (`mlx-community` safetensors and GGUF adapters), reducing memory bandwidth pressure while preserving mathematical precision and perplexity.

### 2. Capabilities
- **Reasoning & Text Generation**: General reasoning, code generation, summarization, and task decomposition (DeepSeek-R1-Distill, Qwen2.5, Llama-3.2).
- **Multimodal Vision & OCR**: Parsing document layouts, receipts, handwritten notes, and photographs locally using compact Vision-Language Models (SmolVLM, Qwen2-VL).
- **Fluid Voice Mode**: Low-latency voice processing with client-side Voice Activity Detection (VAD) and fast text-to-speech synthesis.
- **Model Management**: In-app model browser allowing users to download any compatible repository from Hugging Face.

### 3. Privacy & Security Architecture
- **Air-Gapped Operation**: Once weights are downloaded, the app operates with zero network requirements.
- **Local Sandboxing**: All user conversations, system prompts, history, and uploaded images are strictly isolated inside the iOS application sandbox.
- **Telemetry**: Zero telemetry, zero usage tracking, and no external API calls.

## Supported Hardware
- **iPhone**: iPhone 15 Pro, iPhone 15 Pro Max, iPhone 16 series, iPhone 16 Pro series running iOS 18.0 or later (minimum 6 GB unified RAM, 8 GB recommended).
- **iPad**: iPad Air and iPad Pro models powered by Apple Silicon (M1, M2, M4) running iPadOS 18.0 or later.

## Pricing & Distribution
- **App Store URL**: https://apps.apple.com/us/app/ai-offline-llm-chat-lapis/id6754216427
- **Model**: Completely Free, no recurring subscriptions, no hidden paywalls.
- **Developer**: Álvaro García Pizarro (https://www.lvrpiz.com, hello@lvrpiz.com).