Tencent researchers introduced Gander, an AI model designed to perform simultaneous speech, image, and text processing. The system architecture separates functions into a cerebellum for real-time conversation and a brain for complex reasoning tasks. This design allows users to interrupt the model at any time while it continues to work in the background.
The model segments conversations into one-second intervals to manage responsiveness, maintaining a memory of approximately two minutes. In testing on the Full-Duplex-Bench v3, Gander demonstrated precise timing in all 100 scenarios. It interrupted users in 8 percent of cases, compared to 13.5 percent for GPT-Realtime and nearly 48 percent for other competitors.
Gander was trained on 2.7 million examples, focusing on fluid interaction and silence management. While it excels in timing, researchers noted a tradeoff where the model trails slightly in task accuracy compared to direct text-based inputs. The team plans to release model weights and code via GitHub to support further research in the field.
Tencent continues to expand its AI agent ecosystem, having previously released the Hy3 model in July. The company is currently negotiating a stake in the agent startup Manus. These developments align with Tencent's broader strategy to integrate advanced agent capabilities into existing platforms such as WeChat, WorkBuddy, and Yuanbao.
