Tencent’s Gander separates live conversation from background AI work
Tencent has introduced Gander, a research model designed to converse in real time while completing more demanding tasks in the background. It can process speech, images and text together, continue listening while speaking, and accept interruptions or changes from the user. The system may also ask follow-up questions or report progress without waiting for a prompt.
Gander divides the work between two components. Its “cerebellum” makes rapid decisions about listening, speaking and stopping, using one-second conversation segments and roughly two minutes of recent context. A replaceable “brain” performs reasoning-heavy work such as searching files or fixing code; the report says an unspecified model from OpenAI’s GPT-5.6 family filled that role in testing.
On Full-Duplex-Bench v3, Gander began speaking at the appropriate time in all 100 scenarios and interrupted users in 8 percent of cases, compared with 13.5 percent for GPT-Realtime. However, its task accuracy was slightly below that of the weakest competitor, and it performed poorly on some audio and video-understanding tasks. Tencent plans to release the model weights and training data after its open-source process; a GitHub repository already contains code and demos.
Written by Sakhanda Wire AI from reporting by The Decoder. Facts, quotes and full context are in the original article below.