Tencent introduced Gander, a research model for simultaneous conversation and agent tasks

The Tencent Hunyuan Speech team, together with researchers from several universities, introduced Gander, a model designed to hold fluid conversations in real time while simultaneously processing more complex tasks in the background. According to the technical report, the model processes speech, images and text simultaneously, and users can interrupt it at any time.

The architecture divides tasks between two modules named after human anatomy: the “cerebellum” controls the conversation itself second by second, while the replaceable “brain” handles reasoning and complex agent tasks in the background. According to the report, the brain can be replaced with other agent systems, such as Codex or Claude Code, without retraining the conversational component; in the tests, this role was performed by an unspecified model from the GPT-5.6 family from OpenAI.

In Full-Duplex-Bench v3, Gander had the best timing of all the systems compared, according to the published results – it started speaking at the right moment in all 100 scenarios and interrupted users in only 8 percent of cases, compared with 13.5 percent for GPT-Realtime and almost 48 percent for the weakest competitor (the commercial systems compared also included Gemini Live and Grok). However, the model lagged slightly in task completion accuracy and, according to the authors, performed worse than its base model in understanding video and audio, which they attribute to training focused on conversational fluency at the expense of accurate perception.

The model was trained on approximately 2.7 million examples. Tencent plans to release the weights and training data once the open release process is complete; the code is already available on GitHub, along with sample demos on the project page.