Local Intelligence

Running Large Language Models on Your MacBook with Apple Silicon

Introduction

1. The Problem with Cloud APIs
2. Why Apple Silicon Is Different
3. What This Book Covers
4. Who This Book Is For
5. How to Use This Book

Chapter 1: The Local LLM Revolution

1. Why Local? Privacy, Cost, and Independence
2. The Cloud API Trap: Latency, Censorship, and Hidden Costs
3. Apple Silicon Changes Everything
4. What You Can Actually Run on a MacBook Today
5. How This Book Is Structured

Chapter 2: Apple Silicon Architecture for ML Workloads

1. Unified Memory Architecture Explained
2. CPU, GPU, and Neural Engine: Who Does What?
3. M1 vs M2 vs M3: The Progression for AI Workloads
4. Thermal Design and Sustained Performance
5. Benchmarking Your Mac’s ML Capability

Chapter 3: Model Selection–Finding the Right LLM

1. The Open-Source LLM Landscape in 2025 and Beyond
2. Model Families: Llama, Mistral, Qwen, Gemma, and Others
3. Parameter Count vs Performance: The Sweet Spot for MacBooks
4. GGUF Format and the Quantization Zoo
5. Building Your Personal Model Library

Chapter 4: llama.cpp–The Foundation of Local Inference

1. Installing llama.cpp on macOS
2. Loading Your First GGUF Model
3. Understanding Quantization: Q4_K_M, Q5_K_S, Q8_0, and Beyond
4. Server Mode: REST API and Multi-User Access
5. Advanced Features: LoRA Adapters, Embeddings, and Multimodal

Chapter 5: Ollama–Simplicity at Scale

1. Installing Ollama on macOS
2. The Library: Pulling and Managing Models
3. Modelfiles: Customizing System Prompts and Parameters
4. The Ollama API: Integrating into Applications
5. Creating and Publishing Your Own Models

Chapter 6: MLX–Apple’s Native Machine Learning Framework

1. What Is MLX and Why It Matters
2. Installing and Setting Up the MLX Ecosystem
3. Running Models with mlx-lm
4. Fine-Tuning Models on Your Mac
5. MLX vs llama.cpp vs Ollama: When to Use What

Chapter 7: Hugging Face Transformers on Apple Silicon

1. The Hugging Face Ecosystem on macOS
2. Bitsandbytes and 4-Bit/8-Bit Quantization
3. Hugging Face Optimum for Apple Silicon
4. PEFT: LoRA, QLoRA, and Parameter-Efficient Tuning
5. Building a Complete Local Training Pipeline

Chapter 8: Performance Optimization and Tuning

1. Memory Management: Unified Memory as Both Blessing and Constraint
2. Context Window Optimization and KV Cache Strategies
3. Batch Size, Prompt Processing, and Token Generation Speed
4. Thermal Throttling: Keeping Your Mac Cool Under Load
5. Profiling Tools and Performance Metrics

Chapter 9: Building Local Applications with LLMs

1. The Local Application Stack
2. Chat Interfaces and Web UIs: Open WebUI, Text Generation WebUI
3. RAG Pipelines: Local Retrieval-Augmented Generation
4. Tool Use and Function Calling with Local Models
5. Building a Production-Grade Local AI Assistant

Chapter 10: Fine-Tuning and Customization at Home

1. When to Fine-Tune vs When to Use Prompt Engineering
2. Data Preparation for Local Fine-Tuning
3. QLoRA Training on Apple Silicon with MLX
4. Evaluating Your Fine-Tuned Model
5. Deploying Custom Models in Production

Chapter 11: The Broader Ecosystem–Tools and Integrations

1. Model Serving: vLLM, Text Generation Inference, and Local Alternatives
2. Evaluation Frameworks: Measuring Your Model’s Quality
3. Embeddings Locally: Sentence Transformers and Beyond
4. Multimodal Models on Apple Silicon
5. CI/CD for Local LLM Deployments

Chapter 12: Troubleshooting and Real-World Gotchas

1. Out of Memory: Diagnosing and Solving RAM Exhaustion
2. Slow Inference: Identifying Bottlenecks
3. Model Compatibility and Format Conversion
4. macOS-Specific Issues and Kernel Panics
5. Recovery Strategies and Safe Shutdown

Conclusion: The Future of Personal AI

1. What We Have Learned
2. The Trajectory: Smaller Models, Bigger Chips
3. The Philosophical Shift: From Cloud Dependency to Personal Sovereignty
4. Your Next Steps

References