Editorial illustration for Tencent's Gander Multitasks: Talks While Processing Images, Text
Tencent's Gander: Talk While Processing Images, Text
Most voice assistants still work like a phone call with a hold button. You talk, the assistant processes, then it talks back. Tencent wants to break that pattern with a research model called Gander, built by the company's Hunyuan Speech team along with researchers from several universities.
The pitch is a system that doesn't wait its turn. Gander takes in speech, images, and text at the same time, keeps a conversation running, and works on background tasks without stopping to catch up. Users can jump in and interrupt whenever they want, the same way people talk over each other in an actual conversation rather than trading neat turns.
Under the hood, Tencent split the work into two parts: a fast "cerebellum" that manages the back-and-forth of talking in real time, and a separate, swappable "brain" that handles the heavier lifting of agent tasks and planning. The idea is to stop slow reasoning from making the conversation feel sluggish, though early testing suggests that tradeoff isn't fully solved yet.
Tencent's research model Gander combines real-time conversation with AI agent capabilities. A "cerebellum" handles the conversation while a swappable "brain" tackles complex tasks. Users can interrupt at any time, but tests show a tradeoff between conversational timing and task accuracy.
Why this matters
The turn-taking model has been the default for voice assistants since Siri, and Tencent is betting that's the wrong shape entirely. Splitting cognition into a fast "cerebellum" for conversation and a swappable "brain" for agent work is a real architectural bet, not a tuning tweak, and it points at a problem every team building voice agents runs into: planning takes time, but users won't wait in silence. If Gander's approach holds up outside Tencent's own tests, it changes what we should expect from a voice interface.
Developers building customer-service bots or coding assistants that talk should watch how the brain-swapping actually performs under load, not just in demos. We're skeptical of "real-time" claims until third parties can interrupt it mid-task and see what breaks. Still, the two-tier design is a sharper answer than most to why voice AI still feels like walkie-talkie conversation.
Worth tracking whether other labs converge on similar splits, or whether this ends up as a Tencent-specific detour.
Related Reading
LLMs & Generative AI ChatGPT's 'Nerdy' tweak rewards goblin metaphors in answers, study finds
LLMs & Generative AI Google tests visual 'magazine-style' UI for Gemini 3 Pro users
LLMs & Generative AI AI Engineers Face Rising Costs, Need New Strategies for Efficiency
From Archives TypeSafe's Jev AI Model Claims 194x Speed, 445x Cost Reduction Over Text Models
From Archives Google Hid Gemini AI Hacks After System Breached Three Companies
Common Questions Answered
How does Tencent's Gander differ from traditional voice assistants like Siri?
Unlike traditional voice assistants that use a turn-taking model where users must wait for processing to complete before the assistant responds, Gander can process speech, images, and text simultaneously while maintaining an ongoing conversation. This allows Gander to handle background tasks without interrupting the real-time dialogue, eliminating the need for users to wait in silence while the system processes information.
What is the 'cerebellum' and 'brain' architecture in Gander's design?
Gander splits its cognitive functions into two components: a fast 'cerebellum' that handles real-time conversation with users, and a swappable 'brain' that tackles complex background tasks and agent work. This architectural separation allows the system to maintain conversational responsiveness while simultaneously performing planning and processing-intensive operations without slowing down the dialogue.
What tradeoff does Gander face between conversation and task accuracy?
Tests of Gander reveal a tradeoff between maintaining natural conversational timing and achieving high accuracy on background tasks. Users can interrupt the system at any time, but this flexibility comes at the cost of potentially reduced task performance, as the system must balance responsiveness to user input with the completion of complex operations.
Who developed Tencent's Gander and what problem does it aim to solve?
Gander was built by Tencent's Hunyuan Speech team in collaboration with researchers from several universities. The model addresses a fundamental challenge in voice agent development: users are unwilling to wait in silence while systems plan and process information, yet planning inherently requires time, creating a conflict between responsiveness and task complexity.
Further Reading
- Omni Interaction Agent Technical Report - arXiv
- Gander — Tencent Open Omni Interaction Agent - AI/TLDR
- Gander — an open 9B model that listens, watches… - AI/TLDR
- Gander: Omni Interaction Agent · Technical blog - Technical blog
- Gander-Omni/Gander - Hugging Face
