Smart Home Voice Control That Never Phones Home

What it does Home LLM is a Home Assistant custom component that pairs your smart home with a local language model. It translates chat or voice commands into service calls for lights, locks, climate, and more, and ships with its own family of models—from a 270M-parameter Gemma variant up to a 3B Llama—that are fine-tuned specifically for home automation. If you already run a different model, the integration can teach it your house layout on the fly through in-context learning.

The interesting bit The project treats cloud APIs as a fallback, not a foundation. You can run everything on a Raspberry Pi with no GPU, but if your hardware groans under the load, you can still wire in Ollama, OpenAI, or Anthropic backends. The maintainers also publish their training dataset and scripts, so you’re not stuck with their model choices if your smart home vocabulary drifts from the mainstream.

Key highlights

• Runs fully offline by default; only leaves your network if you explicitly configure an external backend.
• Supports seven different backends including built-in Llama.cpp, Ollama, and generic OpenAI/Anthropic APIs.
• Custom models top out at 3B parameters, with a 270M option for truly modest hardware.
• Controls lights, switches, fans, locks, climate, media players, vacuums, timers, todo lists, and scripts.
• Includes multilingual support for English, German, French, Spanish, and Polish, though the README notes some translations need work.

Caveats

• Routing commands through external APIs exposes the status of every visible Home Assistant entity—including potentially your location—to that provider.
• Some non-English models are explicitly labeled as experimental, and the maintainers are actively seeking better translations.

Verdict Worth exploring if you run Home Assistant and want local, natural-language control without a cloud subscription. Probably overkill if you already have a reliable voice pipeline or prefer not to manage a local LLM backend.