Tencent's Gander interrupts users in just 8 percent of cases, beats GPT-Realtime

Gander reflects a broader push to divide agent work across multiple models. The article notes that OpenAI's GPT-Live separates conversation from reasoning, Sakana AI's Fugu is a separate language model that calls others from an expandable pool, and OpenAI is testing proactive agents that create follow-up tasks. Tencent's own July release of Hy3, an open language model that reportedly narrowed the gap with rivals especially on agent tasks, already runs in WorkBuddy, Yuanbao, and WeChat — and the company is negotiating to take the largest stake in agent startup Manus after Beijing blocked Meta's acquisition. That deal is seen as a fit for its plans, including an agent embedded in WeChat.

Gander was trained on about 2.7 million examples, some of which teach it to stay quiet when there is background noise or nobody in a group is addressing it. The researchers say the work is still early: scaling it up remains an open question, and there is no standard way to evaluate systems like it. Handling interruptions and avoiding delays are practical concerns — an Anthropic analysis found experienced users interrupt Claude Code in about 9 percent of work steps, compared with roughly 5 percent for newcomers.

Whether Gander becomes more than a research showcase hinges on how well it scales and whether its task accuracy improves once the underlying brain model does, since a stronger brain should lift the whole system. For teams building conversational agents, the experiment may offer evidence that treating conversation and reasoning as separate problems is worth the complexity — but only if the accuracy gap narrows in practice.