Home Assistant setup weighs local Qwen2.5

The instinct is right: for Home Assistant, tool use and latency matter more than “chatty” model quality. Qwen2.5-14B can be a good ceiling on this class of hardware, but it is probably not the safest default if responsiveness is the main goal.

• –Home Assistant’s Ollama integration is explicitly designed for local LLMs, and its docs warn that smaller models make more mistakes when controlling the house.
• –A 14B Q4 model can fit the spirit of a 12GB 3060 build, but the margin gets tight once you add conversation history, longer context, and tool-calling overhead.
• –For smart-home use, structured command following beats open-ended reasoning, so a well-prompted 7B/8B model may feel better in practice than a larger but slower one.
• –The best setup is likely a narrow Assist surface with a small set of exposed entities, not a broad “know everything about my home” prompt.
• –If the goal is “more brain than Alexa,” local first is the right architecture; the remaining tuning problem is model size versus real-time usability.