腾讯推出AI模型Gander 可实时对话同步处理后台任务

This article outlines the core features and current development status of Gander, a new AI research model launched by Tencent’s team in September 2026, alongside the broader state of the conversational AI sector, offering readers a quick overview of where next-generation voice assistants are headed.

1. Unlike the turn-taking voice assistants most people use today, Gander can process complex tasks in the background while holding real-time conversations with users. It supports seamless user interruption, and can proactively ask follow-up questions or share task progress updates. In public testing, it delivered the industry’s strongest performance in conversation timing: it interrupted users only 8% of the time, a rate far lower than most competing products.

2. The project’s code is already available on GitHub alongside a live demo, so interested users can test the model directly online. Tencent plans to release the model weights and training data once the full open-source process is complete. The company’s previous generation of comparable AI models has already been integrated into widely used consumer products including WeChat and Yuanbao.

3. This category of technology remains in its early stages. Current limitations include slightly lower task accuracy and weak performance in image counting and localization. Response latency and poor interruption handling remain industry-wide pain points, meaning the technology is still some distance away from large-scale, mature commercial deployment.

The advances in full-duplex AI technology and user behavior insights outlined in this article provide clear guidance for brands planning intelligent interactive services and optimizing user experience.

1. From a user behavior and consumer trend perspective, interrupting mid-interaction is a very common action when users engage with conversational intelligent services: power users interrupt 9% of the time, while new users do so around 5% of the time. Traditional turn-taking voice assistants can no longer fit real-world conversation scenarios. A low-interruption, smooth interactive experience that supports instant interruption and processes tasks concurrently during dialogue is the core user demand for intelligent services.

2. From a product R&D perspective, the dual-module design that separates conversation flow control from complex task reasoning has become the dominant industry technical path. This structure delivers sub-second response speeds while allowing flexible upgrades to backend large language models. At present, widespread industry pain points include response latency and poor interruption experience, with precise audio and video perception and task accuracy as key capabilities to strengthen.

3. For deployment reference, Tencent’s comparable AI technology has already been rolled out in consumer-facing products including WeChat and Yuanbao. Brands can leverage this mature technology to build intelligent customer service and smart shopping guide systems, cutting user wait times for inquiry support.

The development trends in full-duplex AI technology covered in this article offer practical guidance for sellers across platforms to improve operational efficiency and identify new business opportunities.

1. From a demand shift and opportunity perspective, the traditional intelligent customer service and reception tools most merchants use today generally suffer from slow response times and an inability to support users’ need to interrupt mid-conversation. There is clear market demand for a new generation of intelligent reception tools that can handle real-time dialogue while processing backend tasks including order queries, product recommendations and after-sales registration in parallel. Sellers operating in the WeChat ecosystem in particular should closely follow updates on Tencent’s open intelligent agent capabilities, as early integration can deliver a tangible experience advantage.

2. From a reusable experience perspective, when building proprietary intelligent reception systems, sellers can adopt the dual-module split design: use a lightweight frontend module to guarantee sub-second responses and handle user interruptions, and connect backend models with different capabilities to process complex operational tasks. This structure allows flexible service capability upgrades without retraining the frontend module.

3. As a risk note, this category of dual-module AI systems remains in an early development stage, with shortcomings in task accuracy and audio/video recognition, and no unified industry evaluation standard. Sellers should not blindly chase novelty when selecting tools, and instead prioritize testing task completion accuracy in real operating scenarios.

The development trends of next-generation full-duplex AI technology covered in this article provide direction for factories developing smart hardware, advancing digital transformation, and identifying new commercial opportunities.

1. From a product design and manufacturing demand perspective, traditional smart speakers, smart home devices and in-car voice systems equipped with turn-taking voice assistants can no longer adapt to real usage scenarios where users give commands while talking and interrupt interactions freely. Integrating AI systems that support full-duplex interaction and dual-module operation will become the core upgrade direction for next-generation smart hardware, and related hardware manufacturers can pre-develop compatible interfaces to prepare.

2. From a commercial opportunity perspective, full-duplex AI systems are still in a period of rapid technological iteration. The backend reasoning modules of models such as Gander can be flexibly replaced and connected to different intelligent agent systems, so hardware manufacturers can engage with technical teams early to launch new interactive smart hardware adapted to this technology and capture market share. There is also clear cooperation potential for scenario testing and pilot deployment services, as model training requires massive volumes of samples covering diverse background noise scenarios.

3. For digital transformation insights, factories can reference the dual-module split structure to build intelligent production management and control systems: use a lightweight frontend module for real-time production flow responses, and a backend module for complex capacity scheduling and quality inspection reasoning, balancing fast response speeds and complex task processing capabilities.

The technical paths for full-duplex conversational AI and common industry pain points outlined in this article offer clear reference for AI service providers and enterprise service vendors to optimize product capabilities and capture market demand.

1. From an industry development trend perspective, the dual-module architecture that separates conversation control from complex task processing has become a shared technical choice for leading global players including Tencent, OpenAI and Sakana AI. The design, which allows flexible replacement of backend reasoning modules with different large models without retraining the conversation module, can significantly reduce technology iteration costs, and is set to become the mainstream technical direction for next-generation conversational agents.

2. From a customer pain point perspective, all development teams building conversational voice and chat agents currently face widespread issues with response latency and poor user interruption handling. Existing products generally suffer from overly high user interruption rates and inaccurate conversation timing judgment. Meanwhile, there is no unified industry evaluation standard for dual-module systems, leaving customers with no clear reference for vendor selection.

3. For solution reference, service providers can adopt the dual-module design: use a lightweight frontend module with sub-second response to manage conversation flow and process interruption detection, paired with a flexibly replaceable backend reasoning module to handle complex tasks. Providers can also build differentiated competitiveness by making targeted improvements to existing models’ shortcomings in audio/video understanding and task accuracy.

The moves by leading enterprises in intelligent agent deployment and common industry problems covered in this article offer practical reference for platforms of all types to build built-in intelligent features and improve their platform ecosystems.

1. From recent industry practices, Tencent is continuing to scale up investment in open-source large models and intelligent agents. Its previously released Hy3 open-source model has already been deployed across its own platform products including WeChat, Yuanbao and WorkBuddy, and the company is in talks to acquire a majority stake in intelligent agent startup Manus, aligned with its business plan to build built-in agents for WeChat. Its newly launched Gander model has already released open-source code, with model weights and training data set to be opened in the future.

2. From the perspective of real platform user needs, the conversational intelligent assistants currently built into most platforms generally suffer from response latency and poor interruption interaction experiences. Data shows power users interrupt smart assistants 9% of the time, and the traditional turn-taking interaction model can no longer meet users’ demand for natural conversation.

3. For operation and risk mitigation guidance, dual-module full-duplex intelligent agents remain in an early development stage, with no clear large-scale deployment plans and no unified evaluation standard. When rolling out related features, platforms should not blindly pursue larger parameter scales, but instead prioritize polishing interaction fluency and task accuracy. They can quickly fill technical capability gaps and reduce in-house R&D costs through open-source collaboration and investments in startup teams.

This article systematically reviews the latest technical advances, industry deployment landscape and remaining challenges in the full-duplex conversational AI field, providing cutting-edge research directions and industry observation material for researchers in related domains.

1. From new industry trends, the dual-module architecture that splits conversation flow control from complex task reasoning has become a shared technical choice for leading global technology firms. Models including Tencent’s Gander, OpenAI’s GPT-Live and Sakana AI’s Fugu all adopt a similar design, where the flexibly swappable backend reasoning module that supports integration with different large models sharply reduces system iteration costs. Full-duplex interaction, support for seamless interruption, and parallel backend task processing during conversation have emerged as core development directions for next-generation intelligent agents.

2. From unresolved field challenges, existing dual-module systems still face issues including task accuracy that is easily dragged down by errors in the speech recognition link, insufficient precise audio and video perception capabilities, and a lack of scalable large-scale deployment solutions. The industry has not yet formed a unified evaluation standard for dual-module systems, while interaction interruption handling and response latency remain common, unresolved industry-wide problems. Test data shows the rate at which mainstream products interrupt users ranges from 8% to 48%, indicating large experience gaps across offerings.

3. For research reference value, the Gander model was trained on approximately 2.7 million samples, with targeted training samples covering noisy scenarios and utterances not directed at the model to achieve appropriate silent response behavior. Its training approach and “cerebellum + cerebrum” modular design hold notable research value. The project code is now available on GitHub, and its weights and training data will be released for industry research use in the future.