Liu Chuanlin, Chief Solution Architect of Wuwen Xinqiong: Deeply Cultivate the Fertile Ground of Computing Power and Create a Trans - era Super APP | 2025 AI Partner Conference

2025 is the first year of the explosion of AI applications. As the global AI competition enters the "China Moment," a profound technological revolution is quietly reshaping the industrial landscape. At this critical juncture, the industry faces core questions: How can we bridge the gap between AI technology and large-scale applications? Where will the next disruptive AI super-app be born?

On April 18th, the 2025 AI Partner Conference hosted by 36Kr grandly kicked off at the Shanghai Modu Space. The theme of this conference is "The Super App is Coming," focusing on the disruptive transformation of AI applications in all industries. The conference is divided into two major chapters: "The Super App is Coming" and "Who is the Next Super App," covering seven topics such as "Growing Up in the AI World" and "In 2025, Focus on Super Apps in AI." It includes more than 10 keynote speeches, 3 roundtable discussions, and the release of two lists of excellent AI case enterprises, deeply analyzing how AI technology reconstructs business logic and reshapes the industrial landscape, and exploring the infinite possibilities brought by AI super-apps.

On that day, Liu Chuanlin, the Chief Solution Architect of Wuwen Xinqiong, brought a theme sharing titled "Reconstructing the AI Industrial Ecosystem with an Intelligent Foundation."

The following is the content of Liu Chuanlin's speech, edited by 36Kr:

Dear leaders and guests, good afternoon! I am very honored to be invited to participate in this event of 36Kr. As an AI computing power operator and infrastructure builder focusing on supporting AI super-apps, Wuwen Xinqiong's technical capabilities cover the entire chain from computing power optimization, model development to algorithm tuning. We are committed to building a solid technical foundation for the Chinese AI market and helping more innovative applications break through and emerge.

Liu Chuanlin, the Chief Solution Architect of Wuwen Xinqiong

Today's sharing will focus on three cores: First, the trend and transformation of the Infra layer: exploring the technological evolution in the model development, training, and inference stages; Second, Wuwen Xinqiong's practice path: analyzing how we cultivate the growth soil for AI super-apps through technological innovation; Third, future prospects: digging into industry needs and promoting the birth of cross-era AI-native applications.

Looking back at the development history of artificial intelligence, every major breakthrough stems from the upgrade of key elements. Since the emergence of GPT - 3, under the influence of the Scaling Law, both algorithms and computing power have achieved unprecedented development. As the publicly available high - quality text data is gradually exhausted, Ilya predicted last year that the pre - training era is coming to an end.

Taking the GPT series as an example, the iteration cycle from GPT - 4 to GPT - 5 has been significantly lengthened, and the scarcity of pre - training data has become more prominent. The emergence of DeepSeek has brought a new technological paradigm - through R1 reinforcement learning, forming a closed loop of training, inference, and alignment to achieve a second leap in model performance. Its technical path can be summarized as follows: Cold start stage: Based on the R1 Zero model, combined with the Reward Model and alignment algorithm to complete the initial reinforcement; Data optimization stage: Incorporating high - quality industry data and general datasets to improve the model's generalization ability; Closed - loop iteration: Continuously optimizing the model performance through the cycle of "training - inference - alignment - retraining".

This transformation poses dual challenges to Infra: Underlying support: It is necessary to build an AI Infra system suitable for reinforcement learning to meet the needs of large model development; Application empowerment: Through Infra optimization, helping AI applications achieve lower latency, higher efficiency, and better cost, and improving the commercial value (ROI).

As an enterprise incubated by the Department of Electronic Engineering of Tsinghua University, we rely on the technical strength of software and hardware joint optimization to build a technological ecosystem that runs through the upstream and downstream: Computing power layer: Integrating multiple domestic chips to provide diversified computing power support through heterogeneous computing; Platform layer (PaaS): Building an efficient and easy - to - use computing power management platform to improve resource scheduling efficiency; Service layer (MaaS): Providing stable model - as - a - service to lower the threshold of application development.

In cloud services, we adhere to the "trinity" strategy: Multi - heterogeneous adaptation: Compatible with chips of different architectures to ensure flexible supply of computing power; Software and hardware collaborative optimization: Deeply integrating hardware and software to unleash the potential of computing power; Service efficiency improvement: Ensuring efficient use of resources through intelligent scheduling.

Affected by the Sino - US technological game and the rise of domestic chips, in the next three years, domestic chips will become an important carrier for large model training and inference. Therefore, we have launched kilocalorie - level heterogeneous mixed training tasks in Shanghai and other places to overcome the compatibility problems of domestic chips and build a complete ecosystem of "domestic computing power + domestic applications". Through a unified scheduling framework, we can achieve collaborative computing between different chips and significantly improve training efficiency.

In response to the computing paradigm transformation triggered by DeepSeek, we optimize the engineering architecture from three dimensions: Training framework innovation: Self - developing training frameworks suitable for LLM (language model) and MOE (mixture of experts model) to support higher - performance training and be compatible with multiple types of accelerator cards; Communication efficiency optimization: Through the deep overlap of computing and communication, reducing data transmission latency and reducing resource occupation in MOE model training; Dynamic resource allocation: Based on the characteristics of the MOE model, realizing intelligent scheduling of expert models and PD separation to improve the flexibility of the overall architecture.

Taking Shengshu Technology as an example, we provide one - stop services for its multi - modal model training: Second - level environment startup: Quickly deploying the training environment to shorten the project cycle; Automatic fault tolerance: Real - time monitoring of the training process and automatic handling of abnormal situations; Inference efficiency optimization: Through underlying acceleration, improving the response speed after model deployment.

Currently, we have launched a computing power ecosystem service platform at the Shanghai Modu Space and built a benchmark project for government - enterprise cooperation in Zhejiang Province. Through technical support, we reduce the innovation costs of AI enterprises.

In AI application scenarios, inference efficiency directly affects user experience and commercial value. We have launched customized solutions for large language models and text - to - image models: Taking DeepSeek R1 as an example, its MOE architecture and FP8 precision have extremely high requirements for deployment resources. We achieve efficiency breakthroughs through the following optimizations: Engineering transformation: Reconstructing the service framework to increase the inference speed to 30 Tokens/second; Stability guarantee: Optimizing the scheduling algorithm to ensure zero interruption of services in high - concurrency scenarios; Effect alignment: Maintaining the same generation accuracy as the official model while accelerating. Users can call the R1 service with one click through our platform, greatly reducing the deployment threshold.

In response to the traffic fluctuation problem of AIGC applications, we have built an interface - based service based on ComfyUI. Peak shaving and valley filling: Dynamically allocating computing power resources to avoid resource waste during low - traffic periods; Architecture decoupling: Compatibly supporting new models and frameworks through standardized interfaces to reduce repeated development by engineering teams; Multi - modal support: Integrating multi - modal capabilities such as video, voice, and images to achieve seamless docking of the entire industrial chain. Actual cases show that after an e - commerce image generation enterprise adopted our service, it reduced costs, improved inference speed, and enhanced service efficiency.

In the rapid iteration of AI technology, only by grasping the balance between "change" and "unchange" can we seize the opportunity: Change: The technical architecture, computing paradigm, and model form are constantly evolving; Unchange: The core needs of users for efficient, intelligent, and personalized services always exist.

We look forward to in - depth exchanges with experts from various industries, digging into the needs of real scenarios, and jointly promoting the innovation and breakthrough of AI - native applications. As shown by the guests sharing today, whether it is e - commerce product photography, education empowerment, or embodied intelligence, every niche area has the potential to give birth to super - apps. We are willing to use computing power infrastructure as a boat and join hands with partners to sail into the deep waters of the AI industry and create cross - era AI - native APPs.