Tencent unveils Gander to chat during background tasks
Tencent introduced Gander, an AI model that splits real-time conversation and background task execution to let users interrupt and chat with agents while they work.
Tencent's Hunyuan Speech team, alongside university researchers, has developed Gander, an AI model capable of processing speech, images, and text simultaneously. To solve the latency trade-off between quick conversational responses and slow background reasoning, Gander splits its architecture. A cerebellum manages the immediate, second-by-second conversation, while a swappable brain handles complex agent tasks. This brain can be exchanged for external systems like Codex, Claude Code, or models from OpenAI's GPT-5.6 family without needing to retrain the conversational component.
In evaluations, Gander demonstrated superior conversational timing. Tested on Full-Duplex-Bench v3 across 100 scenarios, Gander initiated speech at the correct moment in every instance. It interrupted users in only 8 percent of cases, outperforming GPT-Realtime at 13.5 percent and the weakest competitor at nearly 48 percent. However, Gander lagged slightly in overall task accuracy, partly because speech recognition errors impacted its score. It also underperformed its base model in video and audio perception tasks, such as counting and locating objects, because its training prioritized conversational fluidity over precise perception.
For practitioners, Gander represents a shift toward highly interactive, full-duplex AI agents that can be interrupted naturally. The model processes conversations in one-second segments and retains the last two minutes of dialogue as memory, eliminating the need for a separate voice activity detection module. Tencent trained the model on approximately 2.7 million examples, teaching it to ignore background noise and recognize when it is not being addressed.
Tencent plans to release Gander's weights and training data on GitHub once its open-source review is complete. This release follows Tencent's Hy3 model and aligns with a broader industry trend of delegating tasks across specialized models, similar to OpenAI's GPT-Live and Sakana AI's Fugu. As developers build voice-enabled coding assistants, managing interruptions remains a critical hurdle; for instance, Anthropic found that experienced users interrupt Claude Code in 9 percent of work steps.
This is our own summary of reporting by The Decoder
