Tencent's Gander splits a voice agent into a fast 'cerebellum' for conversation and a swappable 'brain' for background tasks
The research model interrupted users less often than rivals in tests but trailed on task accuracy, and the team plans to release weights and training data.
Tencent's Hunyuan Speech team, working with researchers at several universities, has introduced Gander, a research model designed to hold a real-time conversation while carrying out complex agent tasks in the background. According to its technical report, the model takes in speech, images and text at the same time, and users can interrupt it at any point. The design separates two jobs that current voice assistants usually handle in one place. A component the team calls the cerebellum manages the conversation second by second, deciding when to speak and when to yield. A separate, swappable brain handles longer agent tasks. The split is intended to keep spoken responses fast without forcing the planning component to cut its work short, a tradeoff that today's voice models struggle with because a long-running task can stall the conversation. In tests, Gander interrupted users less often than competing models, but it trailed on task accuracy and showed weaknesses in video and audio understanding. The Decoder characterises this as a tradeoff between conversational timing and task quality rather than a clear win. The team says it plans to release model weights and training data, and a GitHub repository for the code already exists. The work matters for anyone building voice agents that must act as well as talk, since it offers a concrete architecture and early evidence about what the split costs.