Tencent's Gander model uses a dual-system architecture to maintain real-time conversation while performing complex background tasks.
Key facts
- •Gander processes conversations in one-second segments to manage listening, speaking, and interruptions.
- •The model uses the previous two minutes of conversation as memory to inform its responses.
- •Gander interrupted users in 8 percent of test cases, compared to 13.5 percent for GPT-Realtime.
- •Tencent previously released the Hy3 language model in July, which is currently used in WeChat and Yuanbao.
- •The research team is still investigating methods to scale the Gander system.
Researchers from Tencent's Hunyuan Speech team and several universities have introduced Gander, an AI model designed to process speech, images, and text simultaneously. The system aims to mimic human conversation by allowing users to interrupt at any time while the model continues to perform complex tasks in the background.
By the numbers
Dual-System Architecture
Gander splits its processing into two distinct roles: a 'cerebellum' that manages real-time conversation and a 'brain' that handles reasoning and complex tasks. This modular design allows the 'brain' to be swapped for different agent systems, such as Codex or Claude Code, without requiring the conversation model to be retrained.
Performance and Benchmarking
In tests using the Full-Duplex-Bench v3, Gander demonstrated the best timing among tested systems, successfully starting to speak at the correct moment in all 100 scenarios. While it outperformed competitors like GPT-Realtime in managing interruptions, the system's overall task accuracy was slightly lower than the weakest competitor, a result researchers attribute to speech recognition and output errors.
Development and Future Plans
The model was trained on approximately 2.7 million examples, including data to help it remain silent during background noise. Tencent plans to release the model's weights and training data following an open-source release process, and a GitHub repository for the project is already available.
Advertisement
This article was independently rewritten by ManyPress editorial AI from reporting originally published by The Decoder.
