A new Tencent model called Gander is built to hold a conversation and finish a task at the same time. The work comes from the company’s Hunyuan Speech team and collaborators at several universities.
The design premise is that turn-taking is a fiction. Real talk involves interruption, quick acknowledgement and listening during speech, so Gander processes video, speech and text continuously, even mid-sentence, and lets users cut in whenever they like. It can ask follow-ups or report progress unprompted.
One model cannot be both fast and deliberate, so Gander’s designers gave the job to two. Real-time dialogue belongs to what they call the cerebellum. Reasoning and heavier work sit with a separate brain, which can be pulled out and replaced by an agent system such as Codex or Claude Code without touching the conversational half. During tests that slot was filled by an unspecified model from OpenAI’s GPT-5.6 family.
Conversation is chopped into one-second segments, and the cerebellum decides when to listen, speak or yield without a separate voice-activity detector. Roughly the last two minutes serve as memory. On Full-Duplex-Bench v3 the report claims correct onset in all 100 scenarios and an 8 percent interruption rate, against 13.5 percent for GPT-Realtime and nearly 48 percent for the weakest entrant.
Accuracy is where Gander trails. Task completion lands slightly behind, and object counting and spatial questions score worse than the base model, which the researchers blame on training that rewards fluent conversation over precise perception. About 2.7 million examples went into training. No standard evaluation exists for systems like this yet, and weights are promised once an open source review finishes.
