Reachy Mini cuts the cloud entirely with a local AI voice pipeline

A new tutorial from Hugging Face shows how to run Reachy Mini, the open-source robot, completely offline. No cloud. No API keys. No audio data leaving your machine. The setup leans on the team’s ‘speech-to-speech’ library, a cascaded pipeline that chains together voice activity detection, speech-to-text, a large language model, and text-to-speech. The whole thing is exposed through a WebSocket that mimics the Realtime API, which the robot already understands.

For the LLM, they suggest running Google’s Gemma 4 via Hugging Face’s llama.cpp. The recommended flags are practical: two parallel slots so interruptions don’t block the main request, a 64k context window for long chats, and flash attention to keep things snappy. On the voice side, the defaults are Silero VAD, Parakeet-TDT 0.6B v3 for STT, and Qwen3TTS. The blog is refreshingly honest that these are opinionated picks—you’re expected to swap them. New models drop every week, and a cascade architecture means you can hot-swap a faster TTS model or a more accurate STT engine without rearchitecting the whole stack.

What makes this more than a hobbyist flex is the decoupled design. The voice loop and the LLM server talk over HTTP using a Responses API protocol. You can run llama.cpp in one terminal, speech-to-speech in another, and they just… work. This separation means you can benchmark the brain independently from the mouth and ears, which is exactly the kind of unglamorous engineering that separates a demo from something you’d actually leave running on your desk.

There’s a clear tradeoff here that the authors don’t shy away from: “We optimized for multilingual, you might want to optimize for a single language.” No per-minute fees. No vendor lock-in. Just a robot that listens and talks using whatever models you trust. It won’t beat GPT-4o on nuance, but it’s yours. For researchers and tinkerers who’ve been burned by API deprecations or privacy concerns, that’s a real selling point.