By far the most common question I get whenever I talk Home Assistant Voice or LLMs is: “Are you running it locally?” In this video, I can finally say yes—and I walk through my full privacy-first setup: Ollama running on a Mac Mini M4 Pro, integrated into Home Assistant for fast, offline voice control. I show a real-time voice demo (no cloud, no third-party processing), then break down why Apple Silicon is such a compelling local-LLM box thanks to unified memory, and how I’m thinking about model choice as a balance between speed and accuracy. The big takeaway is my “cached responses” strategy. Instead of exposing hundreds of entities and making the model reason from scratch every time, I lean on Home Assistant scripts and predefined pathways. A larger model generates clean summaries (like an hourly spoken weather report), I cache that output, and then a lightweight model can serve it instantly for real-time voice. I also wired up a WLED “SOAP” animation that speeds up based on the Mac Mini’s power draw, so I can literally see when the LLM is working. Finally, I give a quick look at Open WebUI—my local ChatGPT-style interface for testing models side-by-side and running everything outside of Home Assistant.