Gander open-sources full-duplex voice agent: real-time chat front-end with pluggable back-brain

面壁智能 · wechat · 2026-09-17

External researchers led by Shengpeng Ji (Zhejiang University, WavTokenizer author) post-trained MiniCPM-o4.5 into Gander, a fully open-source end-to-end omni-modal interactive agent. A real-time full-duplex "front brain" handles conversation and decides when to delegate to a training-free, pluggable "back brain" (default Codex, swappable with Claude Code) via three tool calls. A chunk-flatten mechanism makes listen/speak/interrupt/tool decisions a per-second model prediction, replacing external VAD. On Full-Duplex-Bench v3 it achieves 100% timely turn-taking with only 8% interruptions vs GPT-Realtime's 13.5%. Task Pass@1 is 0.400 (trailing commercial systems), but text-driven back-brain hits 0.520 with 0.934 tool-selection accuracy, beating all baselines — bottleneck is speech transcription and routing, not architecture. Paper, data, code and weights are all public.

Original post →

More from coding & agent

coding & agent channel →