Tencent Unveils Gander, a Voice AI That Keeps Talking While a Separate 'Brain' Handles Background Tasks

Tencent's Hunyuan Speech team, in collaboration with several university researchers, has introduced Gander, an AI model designed to hold real-time voice conversations while executing complex tasks in the background. According to the technical report published on Arxiv, Gander processes speech, images, and text simultaneously and allows users to interrupt at any point—addressing what the researchers call a fundamental limitation of today's turn-based voice assistants.

A split architecture: "cerebellum" and "brain"

The core design problem, according to the researchers, is that conversation demands fast responses while tasks like searching files or writing code require extended planning time. Rather than forcing one model to balance both, Gander splits the work into two components borrowed from human anatomy: a "cerebellum" that manages the conversation in real time, and a swappable "brain" that handles reasoning and complex agent work in the background.

The brain component is model-agnostic—it can be replaced with systems like Codex or Claude Code without retraining the conversational layer. In Tencent's tests, an unspecified model from OpenAI's GPT-5.6 family filled the brain role, meaning the system's capabilities scale as the underlying reasoning model improves.

The cerebellum processes conversation in one-second segments, deciding whether to listen, speak, or stop when interrupted. Notably, it makes these decisions without a dedicated voice-activity-detection module, instead relying on roughly two minutes of conversational memory.

Benchmark results: better timing, weaker accuracy

Tencent tested Gander on Full-Duplex-Bench v3, a benchmark for voice assistants across varied task scenarios, since no standard benchmark exists yet for systems like Gander. The model started speaking at the correct moment in all 100 test scenarios and interrupted users in just 8 percent of cases—compared to 13.5 percent for GPT-Realtime and nearly 48 percent for the weakest competitor tested, according to the report.

However, Gander's task accuracy fell slightly below that of the weakest competing system, including commercial models such as GPT-Realtime, Gemini Live, and Grok. Tencent attributes part of this gap to how the benchmark scores the full pipeline: speech recognition and output errors count against the system, even though the brain component performs much better when fed text directly.

The model also showed weaknesses in video and audio understanding, underperforming its own base model on some tests—including object counting and spatial localization. The researchers attribute this to training that prioritizes conversational fluidity over precise perception.

Gander was trained on approximately 2.7 million examples, according to the report, including data teaching it to stay silent during background noise or when it isn't being addressed in a group setting.

Open source plans

Tencent says it will release Gander's model weights and training data once it completes "the open source release process." A GitHub repository for the code is already live, with demos posted on the project page. The researchers describe the work as early-stage, noting that scaling the approach remains an open question and that no standardized evaluation method exists yet for duplex conversational-agent systems.

Gander follows Tencent's July release of Hy3, an open language model already deployed in WorkBuddy, Yuanbao, and WeChat. Tencent is also reportedly negotiating for the largest stake in agent startup Manus after Beijing blocked a proposed Meta acquisition.

What this means

Gander is a research preview, not a shipped product—weights aren't public yet, and the accuracy tradeoffs are real. But the architectural bet (decoupling conversational timing from task reasoning) mirrors moves by OpenAI's GPT-Live and Sakana AI's Fugu, suggesting the industry is converging on the idea that one model can't optimize for both fast turn-taking and deep reasoning simultaneously. The interruption-rate numbers are notable, but until Gander is evaluated on tasks scored independently of speech-recognition errors, its actual task competence versus commercial rivals remains unproven.