Gander Kills the Turn-Taking Era

🔥 What happened A 22-author team dropped Gander, a native full-duplex multimodal model built on MiniCPM-o 4.5. It processes streaming input continuously, lets users interrupt mid-sentence, and proactively asks follow-ups. The trick: a Cerebellum-Brain split — Cerebellum for realtime interaction, Brain for heavy reasoning. 💡 Why it matters Every voice agent today waits politely for you to finish — Gander talks and listens at once, like a human. It reportedly handles background noise, multi-party chatter, and backchannel cues. That's the missing piece for agents that actually survive real workflows. ⚡ Our take Turn-taking was a technical limitation, not a law of nature. If Gander delivers, today's voice bots will look like fax machines in the Zoom era.