July's GPT-Live turned "full-duplex speech plus asynchronous reasoning" into a closed product: a conversation layer for listening and speaking, a reasoning layer for getting work done. Two months later the open-source side has its counterpart — on September 9, Tencent's Hunyuan team released Gander, an end-to-end omni interaction agent, with models and code fully open, the paper on arXiv:2609.08977, a #3 spot on the Hugging Face Daily Papers board of the day, and a GitHub repo past a hundred stars on release day.
Two systems: the Cerebellum speaks, the Brain works
Gander's core is a Cerebellum-Brain architecture. The Cerebellum, built on MiniCPM-o 4.5, owns real-time audio-visual perception, full-duplex conversation control, and speech synthesis. The Brain sits behind a pluggable provider interface — Codex by default — and asynchronously handles long-horizon reasoning, tool use, and workflow execution. An agent orchestration runtime connects the two, binding trusted turns, tracking task state, and managing permissions.
Interaction is compressed into one-second causal units: [video][audio][task context] -> [control token][content]. The control token is predicted before content, decoupling whether to act from what to produce into four behaviors: listen, speak, interrupt, tool. The Talker deploys detached from the Thinker, streaming audio at a released alignment of eight text tokens to 50 S3 speech tokens per unit, so perception is never blocked by speech generation; long sessions retain up to 128 recent units. The task side exposes a small vocabulary — task_start, task_send (main/fork), task_resolve — with execution generations fencing stale results after a revised objective. Talk-while-working is thus native: you can interrupt at any time, and it can proactively report progress, ask follow-up questions, or request permission.
The numbers, vendor-reported
Per the technical report, evaluation covers four dimensions — conversational ability, omni understanding, interactive capability, and agentic intelligence — across 2,052 benchmark utterances. The official scorecard: SpokenQA reaches 75.60 / 59.30 on its two subsets and VoiceBench AlpacaEval scores 3.96 with SD-QA at 46.84%, which the team says place first and second in its full-duplex comparison group respectively; Full-Duplex-Bench v3 shows 100% appropriate turn-taking with an 8.0% premature-interruption rate; delegated scenarios bound 45 of 45 final responses correctly; and Daily-Omni holds at 78.53% after full-duplex and agentic post-training. The abstract adds that internal human evaluations confirm natural, expressive speech that stays robust under background noise, multi-party dialogue, and backchannel communication.
The cold water
The repo's dataset badge still reads "coming soon" — of the models-plus-code-plus-data triple, only two have shipped. Every number above is self-reported, with no third-party replication path yet. The default Brain is Codex; swapping in another agent means writing your own provider.
The real increment: the conversation-layer-plus-reasoning-layer split that OpenAI validated inside a closed product now has an open implementation you can take apart — the control-token protocol and the runtime's task lifecycle are all white-box. When voice interaction shifts from "your turn to speak" to "interrupt anytime", the shape of agents changes with it. Is your workflow ready to be interrupted at any moment?