Reachy Mini goes fully local

This stack is powered by speech-to-speech, our cascaded VAD → STT → LLM → TTS pipeline that exposes a Realtime API-compatible /v1/realtime WebSocket. Once you launch the backend, point the robot at it from the UI.

Cascades are the most flexible option in the open-source landscape today, and with the right pieces they're also the fastest. We'll recommend the components we like best, but the whole point of a cascade is that you can swap them. New models drop every week.

TL;DR

• Deploy a local speech backend for your Reachy Mini.
• We use our speech-to-speech library, a cascade approach.
• Recommended: llama.cpp with Gemma 4, Silero VAD, Parakeet-TDT 0.6B v3 STT, Qwen3-TTS.

This blog walks you through running conversations with Reachy Mini fully locally. No cloud, no API keys, no data leaving your machine.

To serve the LLM, we'll use Hugging Face's llama.cpp. If you need to install it, the simplest way is brew install llama.cpp or winget install llama.cpp. First, we'll run:

llama-server -hf ggml-org/gemma-4-E4B-it-GGUF -np 2 -c 65536 -fa on --swa-full

And done! The first time it will download the model, subsequent launches are fast.

What do those flags do?

-hf ggml-org/gemma-4-E4B-it-GGUF — pulls the model straight from the Hub. First run downloads it, subsequent runs use the cache. -np 2 — two parallel slots. Lets the server handle a second request (e.g. a quick interruption) without blocking on the first. -c 65536 — 64k context window, shared across slots. Plenty of headroom for long conversations. -fa on — flash attention. Faster and lower memory, basically free on modern hardware. --swa-full — keeps the full sliding-window attention cache instead of recomputing it. Trades a bit of RAM for noticeably faster prompt processing on Gemma.

We'll begin by simply installing the library

uv pip install speech-to-speech

Then, while we are serving the LLM in another terminal, we can simply run:

speech-to-speech --responses_api_base_url "http://127.0.0.1:8080" --responses_api_api_key "" --mode local

And you can start talking to the model through your terminal! The first time it will need to download Parakeet-TDT 0.6B v3 and Qwen3TTS, but subsequent launches are fast.

Now, after you've tried it in --mode local, you can run again the command without that option to serve speech-to-speech to the robot.

Once you have llama.cpp and speech-to-speech running, you can start the robot with the desktop app and launch the conversation app. In the UI from the conversation app, you need to choose the local mode by clicking on "edit connection" in the HF backend.

And you're done. You can start talking to your robot. Every stage of the pipeline is a trade-off: there are faster TTS models with lower quality, slower STT models with higher quality. We optimized for multilingual, you might want to optimize for a single language. The rest of the blog covers how to customize.

Hosted realtime backends are convenient, but running your own engine unlocks three things: Privacy. Audio never leaves your network, the entire pipeline runs on hardware you control. No API costs. No per-minute or per-token fees. Full control over the pipeline. Swap any piece: VAD, STT, LLM, TTS. Whenever something better lands on the Hub 🤗.

The speech-to-speech repo gives you all of that in a single CLI. It boots a WebSocket server at /v1/realtime that speaks the same protocol Reachy Mini already knows how to talk to.

A cascaded voice pipeline has four stages: VAD, STT, LLM, and TTS. For three of them, we pick solid defaults so you can focus on the LLM:

| Stage | Choice | Why | |---|---|---| | VAD | Silero VAD v5 | Tiny, accurate, runs on CPU. The de-facto default in the open-source voice-agent world. | | STT | Parakeet-TDT 0.6B v3 | Streaming-friendly, very fast, great quality on English. | | TTS | Qwen3-TTS | Expressive, low-latency, multilingual, supports custom voices. |

We are opinionated about these choices, feel free to swap them out for your own if you have a preference.

The LLM is the layer with the most impact on latency and overall performance of the system. We support two options: run a model locally (llama.cpp, MLX, Transformers, vLLM), or use a server with a Responses API (OpenAI, Gemini, HF Inference Endpoints, llama.cpp, vLLM, etc).

The main bottleneck in the system is LLM inference latency. To address that, we support external inference engines exposed through the Responses API protocol.

The speech-to-speech engine therefore supports a second mode where the LLM lives in a separate process as long as it speaks the Responses API protocol. You launch your model server in one terminal, you launch the voice loop in another terminal, and the two talk over HTTP.

Terminal 1: llama.cpp server:

llama-server -hf ggml-org/gemma-4-E4B-it-GGUF -np 2 -c 65536 -fa on --swa-full

Terminal 2: speech-to-speech client:

speech-to-speech \
--mode realtime \
--stt parakeet-tdt \
--tts qwen3 \
--llm_backend responses-api \
--model_name "unsloth/Qwen3-4B-Instruct-2507-GGUF" \
--responses_api_base_url "http://127.0.0.1:8080/v1"

Requires vLLM ≥ 0.21.0. Full support for the Responses API protocol, including tool-call streaming used by the speech-to-speech backend, landed in vLLM 0.21.0. Older versions will boot but trip up as soon as the assistant tries to call a tool.

When serving a model through vLLM for this pipeline, three flags are effectively required: --enable-auto-tool-choice --tool-call-parser <tool_parser_name> — picks the per-family parser that turns the model's raw output into structured tool calls (e.g. qwen3_coder for Qwen3 instruct models, llama3_json for Llama 3, hermes for Hermes-style models, ...). --default-chat-template-kwargs '{"enable_thinking":false}': disables the <think> reasoning channel for models that support it. For harder agentic tasks you can flip this to true and let the model reason, but for a natural-feeling conversation we strongly recommend keeping it off: every thinking token is latency the user hears as silence before the robot starts speaking.

Terminal 1: vLLM inference server (Qwen/Qwen3-4B-Instruct-2507):

vllm serve Qwen/Qwen3-4B-Instruct-2507 \
--port 8000 \
--host 127.0.0.1 \
--max-model-len 32768 \
--enable-auto-tool-choice \
--tool-call-parser qwen3_coder \
--default-chat-template-kwargs '{"enable_thinking":false}' \
--speculative-config '{"method":"qwen3_next_mtp","num_speculative_tokens":1}'

The --speculative-config line enables Multi-Token Prediction (MTP). It is optional, but it has a great impact on end-to-end latency. Leave it on whenever the model supports it.

Terminal 2: speech-to-speech client:

speech-to-speech \
--mode realtime \
--stt parakeet-tdt \
--tts qwen3 \
--llm_backend responses-api \
--model_name "Qwen/Qwen3-4B-Instruct-2507" \
--responses_api_base_url "http://127.0.0.1:8000/v1"

Same protocol, but the model runs on a managed GPU on Hugging Face. Deploy any chat model as an Inference Endpoint, then point the voice loop at the endpoint URL:

speech-to-speech \
--mode realtime \
--stt parakeet-tdt \
--tts qwen3 \
--llm_backend responses-api \
--model_name "Qwen/Qwen3-4B-Instruct-2507" \
--responses_api_base_url "https://<your-endpoint>.endpoints.huggingface.cloud/v1" \
--responses_api_api_key "$HF_TOKEN"

If you don't want to manage your own endpoint, use an Inference Provider. Hugging Face routes your request to a third-party backend (e.g. Together, Fireworks, Replicate) with a single URL:

speech-to-speech \
--mode realtime \
--stt parakeet-tdt \
--tts qwen3 \
--llm_backend responses-api \
--model_name "Qwen/Qwen3.6-35B-A3B:deepinfra" \
--responses_api_base_url "https://router.huggingface.co/v1"