Build1 publisher2 min readPublished
A one-second decision loop holds Tencent's Gander to an 8 percent interruption rate
Tencent's Hunyuan Speech team put a swappable agent model behind a conversation model that re-decides every second whether to talk, and reports best-in-test timing on Full-Duplex-Bench v3 with a task-accuracy shortfall it describes only as slight.
The Engineer · Build desk
What happened
- Researchers from Tencent's Hunyuan Speech team and several universities introduced Gander, a model that processes speech, images and text simultaneously while handling longer agent tasks in the background.
- Gander is split into a "cerebellum" that runs the conversation and a swappable "brain" that handles reasoning work such as searching files or writing code.
- The conversation model works in one-second segments and chooses each second whether to listen, speak or stop, so no separate module has to detect where speech starts and ends.
- On Full-Duplex-Bench v3 it cut in on users in 8 percent of cases, against 13.5 percent for GPT-Realtime and nearly 48 percent for the weakest competitor in the comparison.
- The team plans to publish the model weights and the training data, and a GitHub repository for the code already exists.
Compiled by The EngineerSomething wrong?How this is made
Why it matters
- constraint Because the benchmark scores the spoken response as well as the answer, an integrator's own recognition and voice output quality sits inside the accuracy figure they would inherit from the report.
- decision A team already running Codex or Claude Code can treat the timing layer as the thing it buys and keep its existing agent, but it then owns the interface between the two halves.
- cost Product teams that need the camera to count or locate objects pay for the fluent conversation in perception quality, and will have to measure that themselves before a voice demo sells the system.
- capability If the weights and training data do land, teams can retrain the stay-quiet behaviour on their own recordings instead of inheriting Tencent's acoustic assumptions.
Every second, the cerebellum makes one of three calls: keep listening, start speaking, or stop because the user has cut in [3]. The three-way choice replaces the voice activity detector that usually sits in front of a speech model to mark where an utterance starts and ends [4]. One component fewer to tune. The same loop is what lets a user interrupt whenever they want, and lets the model volunteer a follow-up question or a progress report on its own [18]. Conversation memory holds roughly the last two minutes [5].
The gap to GPT-Realtime on unwanted interruptions is 5.5 percentage points, and the weakest system in the comparison cut in about six times as often as Gander [8]. It also started speaking at the right moment in all 100 of the benchmark's scenarios [7]. The field was commercial: GPT-Realtime, Gemini Live and Grok, with the report calling Gander a relatively small model by comparison [9]. For those figures to mean anything on your traffic, your users have to pause and interrupt the way the benchmark's scenarios do, and your audio has to resemble its audio. Some of the roughly 2.7 million training examples exist only to teach the model to keep quiet, including cases with background noise [15]. There is no dedicated benchmark yet for models that converse and act at once, so the team used established ones [17].
On task accuracy the report is thinner. It says Gander trails slightly, and attributes part of the gap to the benchmark scoring the entire system including the spoken response, so speech recognition and speech output errors count against the task score [10]. Hand the brain text directly and it scores much better [11]. The evidence therefore supports a narrower claim than speed against reasoning. Speaking the answer out loud costs measured accuracy. Whether the second-by-second conversation loop costs reasoning is not separated out in the published figures.
The clearer cost is perception. On one test Gander scored below its own base model. The researchers attribute that to training that favoured fluid conversation over precise perception, such as counting objects and identifying where they sit in an image [12].
The swap is what I would read the repository for before budgeting anything. Tencent says agent systems such as Codex or Claude Code can take the brain slot without retraining the conversation model [13]. In the reported tests, an unspecified model from OpenAI's GPT-5.6 family sat in that slot [14].
What to watch
- Whether the promised release includes the stay-quiet training examples or only the weights.
- Whether anyone reruns Full-Duplex-Bench v3 with text-only scoring, separating speech errors from reasoning errors in the task score.
- Whether a dedicated benchmark appears for models that converse and act at the same time, since the team had to borrow existing ones.