Tencent Gander: Agent That Talks and Works in the Background at the Same Time

Tencent published Gander, a model that holds a conversation and runs background tasks at the same time. The system has two components: the "cerebellum" processes speech in one-second steps, and the "brain" handles complex tasks and can be replaced without retraining. On the Full-Duplex-Bench v3 benchmark, Gander interrupted the user in 8% of cases, compared to 13.5% for GPT-Realtime.

Current voice agents typically wait for a pause in the conversation before accepting a new task. Gander is meant to solve exactly that. Task accuracy is currently slightly below competitors; the authors attribute this to speech recognition errors. The code is already open on GitHub, and the weights and data are promised to follow.