Top Large Model Manufacturers Gather at the TUYA Global Developer Summit to Explore How AI is Reshaping the Industry

On April 23, during the first half of the main forum of the 2025 TUYA Global Developer Summit, Tuya Smart (NYSE: TUYA, HKEX: 2391), the global AI cloud platform service provider, hosted a roundtable forum themed "Disruption and Reconstruction: How AI is Reshaping the Industry Ecosystem".

At the roundtable, Dong Xu, General Manager of Tongyi Large Model Business of Alibaba Cloud Intelligence Group; Ray Chen, North Asia AI Product General Manager of Google Cloud; Di Wu, Head of Intelligent Algorithms at Volcano Engine; Yimin Long, GM of Tencent Cloud Audio and Video AIoT; and Max Ke, Vice President of Technology of Tuya Smart, engaged in an in-depth discussion on cutting-edge topics such as multimodal AI interaction and the commercialization of AI agents.

The rapid evolution of large model AI technology is accelerating the emergence of a vast new industrial landscape. Against this backdrop, how AI can reshape the industry ecosystem has become a critical question across sectors. This roundtable brought together top-tier global players in cloud computing and large model development for the first time, sparking meaningful discourse. It not only delivered key insights to the industry but also offered new perspectives to help developers around the world commercialize AI and further unlock its potential.

Below are selected highlights from the discussion:

Max Ke: To begin with, could each of you briefly introduce your organization's cloud computing and large model-related products and services?

Dong Xu: Alibaba's Tongyi series has always embraced open source and openness as core strategies. We've launched the Tongyi Qwen and Tongyi Wan model series. The number of open-source derivative models has now exceeded 100,000-currently the largest open-source model repository in the world.

Ray Chen: Google Cloud offers a full-stack AI service system, covering a wide range of large models and computing resources for scenarios including audio and video, image processing, and real-time voice interaction. Meanwhile, we provide robust infrastructure support for our global clients, all powered by Google Cloud.

Di Wu: Volcano Engine, ByteDance's cloud service platform, is committed to delivering top-tier AI cloud services. Developers can access full-featured APIs from the Doubao large model and leading domestic open-source model DeepSeek, all available on Volcano Engine.

Yimin Long: Tencent Cloud, as a cloud infrastructure provider, offers a range of services spanning audio and video communication, AI-powered audio and video processing, and security acceleration. Our cloud product suite includes MPaaS media services, EdgeOne edge computing, and CPaaS real-time interaction platforms.

AI Strategy and Layout of Large Model Manufacturers

Max Ke: Open source and closed source are often perceived as a contradiction between technological openness and commercial interests. How does Alibaba Cloud balance between the two? And how should enterprises evaluate whether to adopt Alibaba Cloud's open source models or its closed-source services?

Dong Xu: Open source-and more importantly, continuous open source-is a core value we uphold. Tongyi Qwen is committed to being the best open-source large model. Through open source, we are able to strengthen our talent ecosystem, drive revenue growth on the cloud business, and build highly efficient communication channels with developers beyond traditional commercial frameworks.

Max Ke: How do you view the industry trend toward deploying large models on edge devices in the future? If large models are fully run on the edge, what will be the implications for end-side technology and AIoT chip adaptation?

Di Wu: Over the next few years, AIoT chips and edge computing capabilities will undergo exponential growth. These chips must deliver more accurate sensor input, smarter preprocessing of information, and low-power, low-latency intent recognition. On the cloud side, large models will continue to take on high-difficulty, high-intelligence tasks. In terms of end-to-cloud collaboration, edge and AIoT models are indispensable in scenarios involving weak connectivity or high privacy requirements. When online, edge devices can handle front-end preprocessing and seamlessly collaborate with cloud-based large models to execute complex tasks. In the long term, the edge and the cloud will be mutually reinforcing and deeply integrated. The stronger the edge, the richer the real-world data collected, which will in turn fuel the intelligence and evolution of cloud models. Likewise, smarter cloud models will accelerate the adoption of AGI and intelligent hardware by end users-creating a virtuous cycle of mutual growth.

Max Ke: This year's Gemini 2.0 has demonstrated impressive advances in visual segmentation and robotic arm control-especially Gemini Robotics, which can now execute fine motor tasks via natural language commands. What impact do you think this will have on the embodied intelligence or AI robotics industry? Can enterprises enhance their understanding of the physical world through the open-source Gemma 3?

Ray Chen: At Google, the original intent behind designing large models was to pursue the multimodal and long-context approach. We are now seeing that our multimodal capabilities are becoming increasingly robust. Applying multimodal large models to robotics is not only a business model that is easier to commercialize, but also a logical direction for long-term development. Google Cloud's Gemini empowers robots in four key ways: first, enabling robots to generalize perception, evolving them from "experts" to "generalists"; second, transforming how robots interact with the physical world; third, enhancing human-machine collaboration to make robots more intuitive and human-like; and fourth, replacing traditional code-based instruction with natural language commands, significantly improving management efficiency in robotic applications.

Max Ke: Currently, much of our commercial and technological competitiveness still relies heavily on CPU-based infrastructure. As the industry shifts toward GPU-based computing, what challenges and opportunities do you foresee for cloud service providers?

Yimin Long: In the coming GPU-driven era, cloud vendors must build more flexible and scalable GPU architectures to effectively balance throughput and latency. Tencent Cloud will focus on two strategic directions: first, user-oriented solutions-delivering comprehensive, all-scenario audio and video capabilities powered by leading technologies; second, industry-oriented integration-providing end-to-end cloud-edge collaboration services that help developers embed their solutions into the broader intelligent ecosystem.

The Commercial Implementation Path of AI Agents

Max Ke: According to the METR report, the capabilities of agents in handling complex tasks are doubling every seven months-suggesting that the potential application scenarios for AI are entering a phase of exponential growth. AI agents are becoming a key entry point for many enterprises to innovate in the AI business landscape. In your view, how should enterprises select application scenarios, build their capabilities, and accelerate commercialization?

Dong Xu: One critical direction for AI agents is integrating tool usage with reasoning processes, allowing agents to develop the ability to interact with their environments. Moving forward, Tongyi will continue to enhance model reasoning capabilities, strengthen agents' tool invocation and generalization capacities, and support customers in more efficiently expanding the boundaries of AI applications.

Ray Chen: On the development front, the iteration cycle of large models is becoming increasingly rapid. From a strategic perspective, enterprises should prioritize scenarios where the business model is relatively well-defined. For example, in China, the focus can be on human-centric devices such as smartphones, tablets, and wearables. Overseas, the priority should be sectors driven by efficiency-such as telecommunications, finance, and insurance-where AI agents can effectively help reduce costs and increase productivity.

Di Wu: At present, one of the clearest application paths for AI agents lies in information acquisition and basic data processing. These capabilities are already creating meaningful returns for businesses. I also believe that by late 2025 or early 2026, we will see the emergence of agents capable of handling multi-step complex tasks. At that point, reinforcement learning platforms and hands-on experience in such technologies will become increasingly crucial.

Yimin Long: Beyond utility-based scenarios, AI agents also hold long-term potential in emotional companionship and social interaction. Enterprises should also pay close attention to how agents can be integrated into virtual social environments, real-time interactions, media processing, content comprehension and generation, edge node acceleration, and other intelligent applications. These areas offer promising opportunities for future exploration and innovation.

Through this roundtable discussion, it is evident that multimodal interaction is reaching a critical inflection point for large-scale AI adoption, and the reshaping of the industry ecosystem by AI is now an unstoppable trend. However, the true breakthroughs will come from deep collaboration between technology pioneers and industry developers. Tuya looks forward to continuing to work alongside global developers to redefine the core paradigm of next-generation hardware-and together, seize the opportunities of the AI era.