Show HN: Selora – local model for Home Assistant
github.com
Specs: Qwen3 1.7B base model (Q6 quantized~1.6GB) Four Home Assistant-specific LoRA adapters: - Answers - Clarifications - Automations - Commands ~3.5 GB total download size Runs locally via llama.cpp
We chose a Qwen-based architecture because of a paper on Arxiv (link below) which applied a Qwen based model for local LLM configuration, and showed promising results. We took it a step further in application by training LoRA adapters specialized in Home Assistant configuration.
This alpha release ships our base model with four specialist LoRA adapters preloaded for faster response: answers, clarifications, automations, and commands. The Q6 quantized base model is 1.6GB and the adapters are less than 100MB running either on self-hosted llama.cpp, or on Selora Hub devices, where everything is preconfigured and plug-and-play for you.
We started working on this because the existing options for local LLMs in Home Assistant lack knowledge to run anything useful, so users opt for very large LLMs, typically a cloud model that’s too expensive and not optimized for the kind of high-frequency, always-on smart home use we care about. We think there's a need for open-source models that are small and specialized enough to run on the kind of hardware people actually have at home.
Would love any feedback, questions, or ideas. Thanks for checking it out!
Hugging Face: https://huggingface.co/selorahomes/Selora-AI Arxiv: https://arxiv.org/abs/2502.12923 More details: https://selorahomes.com/selora-ai/