// the find
acon96/home-llm
A Home Assistant integration & Model to control your smart home using a Local LLM
A Home Assistant custom component that lets you control your smart home via a local LLM, either using the author's own fine-tuned models (down to 270M parameters) or any OpenAI-compatible backend. Aimed at privacy-conscious HA users who want voice/chat control without sending home state data to cloud services.
- The fine-tuning pipeline is fully open: dataset generation scripts, training configs (axolotl-style YAML), and the HuggingFace dataset are all included, so you can actually reproduce or extend the models rather than just consuming a black box.
- Backend abstraction is genuinely useful — same integration works with llama.cpp embedded in HA, Ollama, vLLM, LM Studio, Anthropic, and OpenAI Responses API, with per-backend config flows rather than one giant settings dump.
- Has a real test suite (pytest, multiple test files covering config flow, conversation agent, migrations, AI task entities) which is uncommon for HACS custom components.
- The 270M FunctionGemma model is a practical answer to the Raspberry Pi use case — most LLM integrations assume you have a GPU somewhere, this one actually ships a model that can run without one.
- The fine-tuned models only cover a fixed set of HA entity types (lights, switches, fans, etc.). Anything outside that list falls back to ICL with a general model, so reliability degrades significantly for less common domains like input_select, number entities, or custom integrations.
- Multi-language support is thin in practice: only English and German have full ICL example sets; French, Spanish, and Polish are marked experimental with visibly smaller pile files, so non-English users will hit worse accuracy with no clear upgrade path.
- Embedding llama.cpp directly inside Home Assistant via a Python wheel is fragile — version pinning means new HA Python versions (the README already mentions a v0.4.7 fix for Python 3.14) will periodically break the built-in backend, and users on non-x86 hardware need to build the wheel themselves.
- There's no formal evaluation harness or benchmark numbers published for the custom models against real HA entity sets, so it's hard to know how well the 3B vs 270M models actually compare on real-world command accuracy before committing to a setup.