Gander: an open full-duplex voice agent that acts while it talks
Apache 2.0, weights on Hugging Face. A streaming Talker keeps talking while an async Brain runs your tools in the background—if you have three GPUs.
Copy markdown
It talks and works at the same time
Gander splits into a "Brain" and a "Cerebellum": a streaming Talker model holds a natural, interruptible conversation while an asynchronous reasoning agent executes long-horizon tasks in the background. The unlock is that it keeps talking to you while the work actually happens—no dead-air "hold on while I run that."
Turn-taking that actually works
On Full-Duplex-Bench it hit 100% appropriate turn-taking with only 8% premature interruptions, and it's built to handle backchannels ("mhm", "right"), background noise, and multi-party rooms. That timing is the exact thing most open voice agents get wrong.
It sees, not just hears
Built on MiniCPM-o 4.5, Gander streams synchronized audio and video, scoring 78.53% on the Daily-Omni audio-visual benchmark. You can point a camera at something and talk about it in real time, not just paste a transcript.
The catch: three GPUs
Running it locally wants Linux, CUDA 12.4, and three separate GPUs—one each for the Thinker, Talker, and ASR—plus a Codex CLI install to drive the Brain. It's powerful, but it's a workstation project, not a laptop one.
Build this weekend: a self-hosted Jarvis
The weights are Apache 2.0 on Hugging Face and it serves behind a local web UI on port 8000. If you have the GPUs, you can stand up a private, always-listening voice-and-vision assistant that runs tools for you—no per-minute API bill, nothing leaving your machine.