Command Palette
Search for a command to run...
Templates
Reusable full configurations. Build one, then apply it to any Domia in seconds.
Everything on one machine — wake word, mic, STT, LLM, TTS and playback all run locally. Say the wake word and it listens and replies, fully offline. The right choice for a single smart speaker.
Secrets are not included — the target keeps its own.
A headless brain — runs STT, LLM, TTS and intents for the mesh, with no mic and no wake word: it does not listen on its own. Other Domias (a thin client, or a thin identity on the same machine) delegate their inference to it automatically. For a machine that powers a home, not a standalone speaker.
Secrets are not included — the target keeps its own.
Ears and voice only — wake word, mic and playback. It captures speech, then delegates STT, LLM and TTS to whatever capable Domia is nearest on the mesh (an Inference hub, or a full Domia on the same machine) — no delegation setup needed. For low-power edge devices.
Secrets are not included — the target keeps its own.
A thin room device (wake word, mic, playback) that controls Home Assistant over MCP — tools run locally while STT, LLM and TTS auto-delegate to a capable Domia on the mesh. After applying, set the provider URL and paste a long-lived token in the Skills section.
Secrets are not included — the target keeps its own.
Full local pipeline tuned for the NVIDIA Jetson Orin Nano — a small 3B model that fits the 8 GB unified memory, with the LLM on the GPU and STT/TTS on CPU. A starting point: the exact 3B model and quantization are still to be confirmed against on-device benchmarks. Pull the model in Ollama first.
Secrets are not included — the target keeps its own.
The Jetson pipeline with hearing moved off the CPU — STT runs against a local NeMo-Speech server (nemotron-3.5-asr-streaming-0.6b) instead of an in-process sherpa model, and the LLM stays on llama.cpp. The most accurate Jetson starting point, but it needs two servers running before it works: start NeMo-Speech with `make nemo-serve` (the URL must end in /v1) and llama-server with `make llm-server`.
Secrets are not included — the target keeps its own.
Standalone tuned for the lowest time-to-first-audio — a fast 3B model (llama3.2:3b) and Parakeet STT (accurate, punctuation + case), with a minimal prompt: no emotion, facts or recent-turn memory injected on the hot path (capture still runs, so nothing is lost). Trades rich context for responsiveness — a snappy ~1s TTFA on a capable machine. Pull llama3.2:3b in Ollama and run `npm run setup:models:parakeet` first.
Secrets are not included — the target keeps its own.
The middle of the quality ladder — fast 3B voice loop (Parakeet + Kokoro, speculation on) WITH the companion layers on: emotion, memory window, facts. Reflection runs on a separate 1B model so it never evicts the voice KV cache. Pull llama3.2:3b and llama3.2:1b in Ollama first.
Secrets are not included — the target keeps its own.
Top of the quality ladder — all companion layers on, wider memory window, reflection on its own 1B model. Uses llama3.2:3b; pull llama3.2:3b and llama3.2:1b in Ollama first.
Secrets are not included — the target keeps its own.
A full standalone Domia that listens, thinks and replies in Spanish — multilingual Whisper STT (language forced to es) and a Kokoro Spanish voice (ef_dora). The wake word stays English ("computer"). Run `npm run setup:models:whisper-multilingual` first; the LLM (qwen2.5:3b) speaks fluent Spanish; pull it in Ollama first.
Secrets are not included — the target keeps its own.