Three False Starts: How I Finally Got Local LLM Working with OpenClaw

From $150/month in cloud subscriptions to $3/month running AI on my own hardware. This is exactly how I got there.

Six weeks. Three failed setups. One breakthrough.

I tried Ollama — too slow for real work. I tried LM Studio — better, but still not production-ready. Then I found llama.cpp paired with Qwen3.5-35B, and everything changed: 3–4x faster inference, zero hallucinations across two weeks of daily use, and a setup that runs overnight automations without throttling, rate limits, or anyone else's terms of service.

This book documents every failure and every breakthrough in precise detail — not as a success story cleaned up after the fact, but as it actually happened.

What's inside:

• Why Ollama and LM Studio fell short — and the specific bottlenecks that killed each attempt
• The llama.cpp breakthrough: exact build flags, model configuration, and the one setting that made the difference
• Complete Mac Mini M4 Pro setup guide: 64GB RAM, llama.cpp, Qwen3.5-35B, and OpenClaw wired together
• Side-by-side performance comparisons: speed, output quality, and real-world cost
• The honest sovereignty assessment — who should go fully local, who shouldn't, and how to decide

Written for OpenClaw users and developers who are done paying for AI they don't own.

Brandon Pham is the founder of xFlowra, an AI automation consultancy, and has been building local-first AI workflows since 2025.