A new open AI agent splits fast talk from slow background thinking
A research paper describes Gander, a 9 billion parameter open agent credited to Tencent's Hunyuan Speech team, that splits real-time conversation from long-horizon reasoning so it never goes silent mid-task.
An open 9B agent splits real-time chat from background reasoning, so it never goes quiet mid-task.
Ask a voice assistant a question that needs real thought, and most go quiet for a few seconds while a model works, then answer all at once. A paper posted to arXiv on September 9, 2026 describes an open agent built to skip that silence entirely. The paper's authors are credited by The Decoder and other coverage as Tencent's Hunyuan Speech team, though the paper's own text does not spell out institutional affiliations as plainly as a typical academic byline, so treat that attribution as reported rather than self-declared.
The agent, called Gander, has 9 billion parameters and is built on top of an existing open model, MiniCPM-o 4.5. Its architecture splits into two parts: a fast component the paper calls the Cerebellum, which handles real-time perception and response in roughly one-second chunks, and a swappable component called the Brain, which handles longer reasoning tasks like searching code or files. The Cerebellum keeps a conversation moving naturally while the Brain works in the background, then hands off what it finds.
Code and demos for Gander are already public on GitHub. The model's weights and training data are not yet released, described in the project's own materials as pending an internal open source review process, so outside researchers cannot yet run or retrain the full model themselves. The project ships under an Apache 2.0 license, inherited from the MiniCPM-o base model it builds on.
What the paper does not yet show is how Gander performs against a real conversational workload outside its own benchmark, or how long a Brain-side task can run before the Cerebellum's small talk starts to feel like stalling rather than conversation.
Why a build studio cares
We build AI workflows and custom agents, and the specific problem Gander is aimed at, an assistant that has nothing to say while it works, is one every voice or chat interface we have shipped has to solve somehow, usually with a canned "let me check that" filler line rather than an architecture built for it. Splitting fast perception from slow reasoning into two swappable components is a genuinely different answer to that problem than a longer filler phrase, and it is worth testing against a real client workload before assuming it holds up outside a benchmark.
Next step: read the paper on arXiv and check the public code on GitHub. If you are building a voice or chat product that goes quiet at the exact moment a user needs it to keep responding, write to us at hello@gattyworks.com.