텐센트 '간더', 작업 수행 중에도 대화 지속하는 AI 모델

Tencent's Gander aims to keep talking while it works in the background Jonathan Kemper View the LinkedIn Profile of Jonathan Kemper Sep 20, 2026 GPT-Image-2 prompted by THE DECODER Key Points Tencent's Gander is designed to hold real-time conversations while handling complex tasks in the background. It processes speech, images, and text simultaneously, and users can interrupt at any time. A "cerebellum" manages the conversation second by second, while a swappable "brain" handles complex agent tasks. The split aims to keep responses fast without cutting planning short. In tests, Gander interrupted users less often than competing models but trailed on task accuracy and showed weaknesses in video and audio understanding. The team plans to release model weights and training data. A GitHub repository for the code already exists. Ask about this article… Search 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. Researchers from Tencent's Hunyuan Speech team and several universities have introduced Gander, an AI model designed to talk, watch, and work at the same time. According to the technical report, it processes speech, images, and text simultaneously, holds real-time conversations, and handles more complex tasks as an agent. The researchers argue that today's voice assistants mostly take turns with users. A user says something, the model responds, and then the user gets another turn. People interrupt each other, give brief feedback, and listen while they speak. Gander continuously takes in video, speech, and text to make those exchanges more fluid. Users can interrupt at any time, while the model can ask follow-up questions on its own or report on its progress. Ad Separating conversation from reasoning keeps the chat responsive The researchers split the model into two roles they call the "cerebellum" and the "brain," borrowing from human anatomy. The cerebellum handles real-time conversation, the brain takes on reasoning and demanding tasks such as searching files or writing code. Ad This split aims to address a tradeoff. Conversation needs quick responses, but complex tasks take time to plan. The researchers argue that a single model inevitably has to balance speed against reasoning ability. Existing agent systems such as Codex or Claude Code can replace the brain without retraining the conversation model. In testing, an unspecified model from OpenAI's GPT-5.6 family served as the brain. When the underlying model improves, the whole system benefits directly. Ad Gander gets the timing right but trails on task accuracy To keep the cerebellum responsive, Gander breaks conversations into one-second segments. Each second, the model decides whether to listen, speak, or stop speaking, for example when the user interrupts. That lets it respond at the right moment without a separate module to detect when speech starts and stops. Its memory holds roughly the last two minutes of conversation. There's no dedicated test for these models yet, so the researchers used established benchmarks. According to the report, Gander has the best timing on Full-Duplex-Bench v3, which tests voice assistants across different task scenarios. Ad It starts speaking at the right moment in all 100 scenarios and interrupts users in 8 percent of cases. The report compares that with 13.5 percent for GPT-Realtime and nearly 48 percent for the weakest competitor. Gander achieves these results with a relatively small model competing against commercial systems including GPT-Realtime, Gemini Live , and Grok, the report says. Ad Gander trails slightly on task accuracy alone. The researchers partly attribute this to the test scoring the entire system, including its spoken response. Errors in speech output and speech recognition also count toward the score. When the brain receives text input directly, it scores much better. The researchers also acknowledge weaknesses in video and audio understanding. In one test, Gander performed worse than its base model. They attribute this to training that prioritizes fluid conversation while neglecting precise perception tasks, such as counting objects and identifying their positions in an image. Tencent plans to release Gander's weights and training data Gander was trained on about 2.7 million examples, according to the report. Some specifically aim to teach the model to stay quiet, such as when there's background noise or nobody in a group is addressing the assistant. The researchers describe Gander as an early exploration of an open research area. How to scale up the model remains unresolved, and there's no standard way to evaluate systems like it. The team plans to release model weights and data. A GitHub repository for the code already exists, and demos are available on the project page . More companies are delegating agent tasks to other models Tencent released Hy3 in July, an open language model that reportedly narrowed the gap with rivals, particularly on agent tasks. It's already used in WorkBuddy, Yuanbao, and WeChat. The company is also negotiating to take the largest stake in agent startup Manus after Beijing blocked Meta's acquisition. Tencent sees the deal as a fit for its own plans, including an agent embedded in WeChat. More companies are building agents around orchestrators that assign tasks to other models. OpenAI separates conversation from reasoning in GPT-Live . A background model handles tasks that need web search or agent capabilities while the conversation continues. Sakana AI's orchestrator Fugu is a separate language model that calls other models from an expandable pool. OpenAI is also testing proactive agents that create their own follow-up tasks and contact users on their own. An Anthropic analysis found that experienced users interrupt Claude Code in about 9 percent of work steps, compared with roughly 5 percent for newcomers. Teams building conversational voice and chat agents report latency problems particularly often, according to a survey . AI News Without the Hype – Curated by Humans Subscribe to THE DECODER for ad-free reading, a weekly AI newsletter, our exclusive "AI Radar" frontier report six times a year, full archive access, and access to our comment section. Subscribe now Source: Arxiv