Tencent's Gander talks while a second model reasons in the background
Tencent's Hunyuan Speech team and five university labs want voice agents to stop taking turns. Their answer splits the talking from the thinking.
Gander runs two models instead of one: a 9B "cerebellum" that handles real-time listening and speech, and a larger "brain" that does the reasoning and tool work behind it. The cerebellum consumes one-second audio chunks and decides at each step whether to listen, speak, or stop — no separate voice-activity detector, no waiting for an acoustic pause to end a turn. It holds roughly the last two minutes of conversation in memory, and when a request needs real work it passes the transcript plus the latest video frames to the brain through a tool call. The whole system was trained on about 2.7 million examples, including deliberately negative ones: unrelated video, background noise, and multi-party rooms where nobody is addressing the assistant.
The benchmark picture is genuinely split, and the authors do not hide it. On turn-taking, Gander takes the floor at the right moment in all 100 evaluated scenarios and speaks prematurely in 8.0% of turns against 13.5% for GPT-Realtime; Gemini Live 3.1 keeps interruptions lower at 19.2% but answers in only 78.0% of scenarios. On task accuracy Gander trails — Pass@1 of 0.400 against 0.600 for the strongest baseline, with tool selection at 0.759 and argument accuracy at 0.503. Run the brain directly from the text transcript, though, and tool selection climbs to 0.934, above GPT-Realtime's 0.876. The execution tier is fine; the delegation step and the speech path are where accuracy leaks.
The part worth watching is what Tencent says it will ship: the weights, the training data, and a public repo. That is a research lab behaving like an open-source project, and it puts a full-duplex multimodal agent within reach of anyone who wants to build on the architecture rather than read about it. The takeaway is the architecture, not the scorecard — a 9B model that keeps a conversation alive while a bigger model grinds on a task is the right shape for voice agents, and judging Gander on Pass@1 alone misses that the paper's own scoring counts speech recognition errors against the composite.
Vercel says Jev is the fastest-adopted model in AI Gateway history — about 13% of teams reached in its first day, twice the GPT-5.6 family and six times Fable 5.1. Cloudflare added the TypeSafe AI decision model to its stack within days as well, per Forbes. Jev returns typed choices, scores, and probabilities instead of prose, and TypeSafe's own workflow evaluations put it up to 194 times faster and 445 times cheaper than language models doing the same judging job. We covered the wave earlier today — APUS reverse-engineers Jev into a fully offline browser agent. Adoption figures from one gateway are not market share, but the direction is clear: the boring infrastructure slots — routing, verification, retry decisions — are being pulled out of the chat model and into something cheap enough to call at every step.
What to watch: whether Gander's weights actually land, and whether the cerebellum-brain split survives scaling. The paper leaves both open.
Does a voice agent that keeps chatting while it works feel more natural to you, or does the filler get old fast? Tell us in the comments.
Sources: Omni Interaction Agent Technical Report (arXiv) · The Decoder · Gander project page · Gander code (GitHub) · Vercel · Forbes · TechCrunch
