Tencent's Gander aims to keep talking while it works in the background
decoder ·
Tencent's Hunyuan Speech team and university researchers have introduced Gander, a model designed to converse in real time while handling complex agent tasks in the background. It processes speech, images, and text together, allowing users to interrupt it, ask follow-up questions, and receive progress updates. Its architecture separates a fast conversational "cerebellum" from a swappable "brain" that performs reasoning and complex tasks, with systems such as Codex or Claude Code able to fill the latter role without retraining the conversation model. On the Full-Duplex-Bench v3, Gander started speaking at the right moment in all 100 scenarios and interrupted users in 8 percent of cases, outperforming the compared systems on timing. However, its task accuracy was slightly below that of the weakest competitor, and it showed weaknesses in video and audio understanding. Gander was trained on about 2.7 million examples, and Tencent plans to release its model weights and training data while a code repository and demos are already available.
decoder