Tencent Gander AI is Tencent’s Hunyuan Speech–led multimodal duplex agent built to keep a real-time conversation going while a separate “brain” runs longer tasks in the background. The Decoder’s Jonathan Kemper summarised the 20 September 2026 research push for a general audience: speech, images, and text in parallel; interruptible turn-taking; a streaming Thinker-Talker “cerebellum”; and benchmarks against GPT-Realtime and Gemini Live. If you are searching how Gander differs from turn-based voice bots—or whether weights are open yet—this TRENDS Technology explainer sticks to The Decoder and the Omni Interaction Agent report (arXiv 2609.08977).

What problem Gander is trying to solve

Most voice assistants still take turns: you speak, they wait, they reply. Real conversations overlap—people interrupt, give quick feedback, and listen while talking. Gander is designed for that full-duplex pattern. It continuously processes video, speech, and text even while it speaks; users can jump in at any time; the system can ask follow-ups or push progress updates without a fresh prompt. Example workloads The Decoder highlights include fixing a bug or waiting for a specific slide to appear while the chat stays live.

Cerebellum vs brain: the anatomy split

Conversation wants sub-second timing; searching files or writing code needs planning time. Gander splits roles: a cerebellum manages second-by-second listen / speak / stop decisions via a streaming Thinker-Talker architecture, while a swappable brain runs asynchronous agent tasks. The brain can be swapped for systems such as Codex or Claude Code without retraining the conversation model; The Decoder notes tests used an unspecified model from OpenAI’s GPT-5.6 family in that slot. As the underlying brain improves, the whole stack benefits.

Gander segments conversation into roughly one-second windows and uses about the last two minutes as memory—without a separate voice-activity-detection module.

People also ask: how does Gander score vs GPT-Realtime?

On Full-Duplex-Bench v3, The Decoder reports Gander started speaking at the right moment in all 100 scenarios and interrupted users in 8% of cases, versus 13.5% for GPT-Realtime and nearly 48% for the weakest competitor tested. Task accuracy trailed slightly; researchers partly blame end-to-end scoring that penalises speech recognition and output errors, noting the brain scores better when given text directly. Video and audio understanding also lagged the base model in one test—training that favours fluid conversation over precise perception, including object counting and localisation.

Open weights, code, and what is shipping

Training used about 2.7 million examples, including cases that teach the model to stay quiet when background noise or group chat is not addressed to it. The team plans to publish weights and training data after completing “the open source release process.” A GitHub repository for code already exists, with demos on the project page; Hugging Face listings under Gander-Omni / Gander track the public artefacts. Treat Gander as research-plus-code today—not a WeChat consumer toggle—while noting Tencent’s parallel agent push (Hy3 in WorkBuddy / Yuanbao / WeChat; Manus stake talks) as company context, not Gander GA.

Dual-geo search: duplex agent vs voice-assistant keywords

US and India queries often split between “Tencent Gander AI” / Omni Interaction Agent paper language and “duplex voice AI / interruptible chat” capability terms. Keep Tencent Gander AI primary; fold Full-Duplex-Bench and Cerebellum/Brain into H2s. Update path: when weights land on Hugging Face, revise the open-source H2 first.

Related Sunday Profile coverage

Deepen Technology explainers with our Runway Real-Time AI Video (https://www.thesundayprofile.com/runway-real-time-ai-video) TRENDS piece and OpenAI Codex Sandbox Escape (https://www.thesundayprofile.com/openai-codex-sandbox-escape) BREAKING—same AI-tools cluster, distinct primary keywords from Tencent’s duplex agent.

Sources: The Decoder, Jonathan Kemper, 20 Sept 2026; arXiv 2609.08977 Omni Interaction Agent; Hugging Face / GitHub Gander project pages.