Google's Gemma team has published a demo of a real-time voice chat with the desktop robot Reachy Mini. Natural conversation runs through the cloud-based Gemini Live API, and the demo ends by running the lightweight Gemma 4 model locally on the robot itself.
Reachy Mini · Open-Source Desktop Robot Demo
One Robot, Two Brains: Cloud Voice Chat or Fully On-Device AI
Google's demo runs the same desktop robot two ways — a real-time voice conversation through the cloud Gemini Live API, then switching to Gemma 4 running locally on the same machine. Social robotics, in the cloud and on the edge.
9
degrees of freedom, incl. Stewart-platform neck
256K
context window in Gemma 4, multimodal input
200+
example apps in the open-source app store
Two hardware editions · price
Gemini Live — Cloud
Runs: CloudStrength: Low latency, vision integrationConnection: API via WebRTCLicense: Gemini API
Gemma 4 — On-Device
Runs: On-deviceStrength: Privacy, no cloud neededConnection: llama.cpp / MLX / WebGPULicense: Apache 2.0
The voice loop
Vision + Audio in
→
Model reasonscloud or local
→
Function call drives actuatorshead, antennae, dance
A fully local Gemma 4 E4B + llama.cpp setup ran so fast that a deliberate delay was added to stop it interrupting — a sign on-device social robotics is entering practical use.
What works
Natural real-time voice, responses to jokes and commands, dance and face tracking. Privacy users value running entirely off the cloud.
Open challenges
"More for experimentation than deployment." Local mode needs stronger hardware; calls remain for better security (no basic auth) and stability.
Continue reading The rest of this article is for AI News Blitz readers. Choose an option below to keep reading.
Already purchased? Sign in ✓ Signed in — this article isn’t included in your current plan.Unlocking the full article…