By Sihai

Source / Node Finance

The landscape of large models has shifted dramatically, from the intense competition among hundreds of models to heated discussions about Product-Market Fit (PMF). DeepSeek's comprehensive free and open-source initiative has sparked a global revolution, not only reducing computational costs but also altering the dynamics between open-source and closed-source models. Virtually all major technology firms, internet companies, and startups are now announcing the open-sourcing of their large models or rebuilding their businesses around open-source frameworks.

If the high costs of closed-source models previously hindered their adoption by enterprises, how can large models now reshape business models amidst the rising tide of open-source alternatives? What constitutes a truly thriving AI ecosystem? What challenges and misconceptions do enterprises encounter when leveraging large models to enhance their operations? We interviewed eight practitioners, spanning B-end technology service providers and C-end application parties, to delve into the essence and application methodologies of large models.

01 Large Models: The Essential Utilities of the AI Era

As the large model ecosystem flourishes, a question arises: What is the killer application of the large model era? Is it Yuanbao, Doubao, or ChatGPT? Amidst significant marketing investments to drive traffic and daily activity, insiders are scrutinizing every detail to find the WeChat or QQ of this new era.

According to , the fundamental difference between large models and the mobile internet lies in their roles: the former as a "productivity tool" and the latter as a "reconstructor of production relations".

An Xiaopeng, Vice President of Alibaba Cloud Intelligence Group, told that large models excel in deeply integrating across thousands of industries, functioning as intelligent brain systems embedded in various hardware devices, collaborating with humans in software to complete tasks, or serving as control systems for autonomous driving, directing other business systems to automate workflows. This diversity enriches their value propositions and business models. If we seek a suitable reference point for large models, integrated circuits align more closely than the internet.

For different application methods, interviewed application-end enterprises. Across our discussions, we found that efficiency gains and cost reductions are the most notable benefits of large models.

Considering control systems, Huang Yan, founder of Leikezhitu, which specializes in providing unmanned solutions for underground mines, noted that Chinese mining machinery brands previously sold at half the price of overseas brands. However, after reconstructing scenario operation capabilities with large models, their unmanned mining machinery now commands prices on par with overseas brands, enhancing operational capabilities and increasing brand premium.

In the context of software collaboration with enterprise businesses, Li Weihan from New BlueFocus Digital Group revealed that the application of agents has significantly boosted content production efficiency. In the first quarter of 2025 alone, AI-driven revenue at BlueFocus approached that of the entire previous year, with this year's estimate reaching a scale of 3 to 4 billion yuan.

Regarding embedding as an intelligent brain in hardware devices to enhance scenario application capabilities, Wang Qi, CEO of Ruiwo Technology, which focuses on intelligent product R&D for hotel scenarios, shared that leveraging the open-source nature of large models, Ruiwo Technology can rapidly develop new functions, reducing the self-service check-in process from 2 minutes to 10 seconds, diverting 60% of guests during peak hours, and saving small and medium-sized hotels at least two front desk labor costs.

While understanding how large models penetrate enterprises or industries is important, exploring how enterprises use them to transform their businesses is even more valuable.

An Xiaopeng believes that to objectively gauge a country's AI industry today, we must focus not only on model rankings but also on the robustness of the entire software application ecosystem. This is the criterion for judging the application and implementation of the AI industry. Li Yanhong has also emphasized that whether open-source or closed-source, model rankings are secondary; what matters is application.

Kunlun Academy, a leading domestic entrepreneurship training institution, shares this perspective. Basic large models require high computational costs and are of concern to a select few enterprises. Most enterprises should prioritize optimizing scenario applications.

The market is the ultimate test of technology value and product competitiveness. So, how do pioneering enterprises form a commercial closed loop?

02 Harnessing Large Models for Application-End Enterprises

Large models penetrate business scenarios through three forms: embedding as intelligent brains in hardware, software collaboration, and automated system control. Unleashing their full potential hinges on enterprises' ability to utilize unique scenario data to form a business closed loop. Kunlun Academy asserts that technology must be deeply integrated with application scenarios to avoid the paradox of possessing a powerful tool but lacking suitable applications.

Wujie Data, whose core business involves public opinion monitoring, used AI to enhance its platform before the rise of large models, but the AI was not sufficiently intelligent. For instance, sentiment analysis of 100 articles using ordinary AI had an accuracy of only 70%. With the advancement of large model technology, Wujie Data's sentiment analysis accuracy for text and video content rose to 98% after fine-tuning its large model with its own data, significantly reducing labor costs for public opinion monitoring.

Tian Wenjun, founder of Wujie Data, believes that large enterprises have extensive operations. Beyond their main reach, grasping business opportunities in niche scenarios and empowering large models with unique data to form a business closed loop is the commercialization path for small and medium-sized enterprises.

Leikezhitu is another example of leveraging data combined with large models to empower business scenarios, thereby enhancing enterprise competitiveness. Huang Yan stated that for mining scenarios, most general large models on the market have excess computing power, but what is truly lacking is data from niche scenarios.

Accumulating data is no easy feat. Mining scenarios focus on non-standard underground transportation scenarios, with far less data than public roads. Moreover, general large models struggle to directly meet the specialized needs of mines, necessitating customized algorithm optimizations aligned with industry characteristics.

Leikezhitu empowers mining L4 unmanned scenario applications through fine-tuned large models, enhancing the perception and efficient coordination capabilities of decision-making modules in complex scenarios, ultimately reducing underground operation manpower by nearly 50% and improving mining scenario safety. Their self-produced mining machinery has gradually achieved technological parity with overseas brands, transitioning from a past situation where "even at half the price of imported equipment, recognition was difficult to obtain".

Beyond deeply exploring niche scenarios to enhance competitiveness, the application of large models by major enterprises is not focused on single-point breakthroughs but rather on improving efficiency across the entire value chain.

Ma Zihan, co-founder of Danghong Qitian, candidly stated that large model empowerment spans the entire company's business, encompassing content creation, user insight analysis, and post-operation optimization.

Danghong Qitian Group is an AI+XR eco-cultural technology enterprise with a full-chain layout to create immersive digital and intelligent experiences. It integrates cutting-edge technologies such as artificial intelligence (AI), 5G cloud rendering, large-space positioning, advanced sensors, and 5G-A, applied in diverse scenarios like technology, culture, tourism, education, and esports. A project's implementation typically involves multiple links, including user market research, content creation, implementation, and user data analysis.

As a non-standard industry, content complexity is exacerbated by differences in user preferences and cultures across regions. Each project's planning has non-replicable characteristics. However, the enhancement in large model capabilities addresses this issue.

Ma Zihan revealed that large models can participate in multiple links such as code development, script optimization, and image generation, shortening the content production cycle. They can also be applied to user insights, helping Danghong Qitian closely grasp regional cultural characteristics and trends. Additionally, XR/VR virtual scenario experiences involve multimodal interaction, and AI can track user experience behavior, such as interaction duration, capture user feedback on content, and use this to inform product iterations.

As large models become nearly standard in vertical niche areas, how can enterprises avoid the "theater effect"? The answer lies in agents.

Li Weihan from New BlueFocus Digital Group views agents as the inevitable evolution from large models to scenario applications. An Xiaopeng echoed this sentiment, stating that agents are the new form of software in the AI era, akin to SaaS a decade ago.

According to observations, agents are playing an increasingly crucial role in company operations. Let's take the content-centric marketing industry as an example.

Since the popularization of AIGC, there has been a deluge of homogeneous content on the market. Producing high-quality content to attract users' attention is a challenge every marketing firm must address. BlueFocus has provided a solution.

As early as 2023, BlueFocus began aggressively deploying full-link AI marketing, collaborating with leading platforms like Alibaba Cloud, Baidu Intelligent Cloud, and Volcano Engine to create a dedicated marketing large model, BlueAi.

However, the vertical marketing model is just the beginning for BlueFocus. "Customers in different fields still face expertise and information barriers," said Li Weihan. To this end, BlueFocus has utilized data and experience accumulated from serving customers to generate agents. Even industry novices can deliver excellent results for brand clients with the help of these agents. Currently, BlueFocus has created over 100 agents, with more than 30 reaching expert levels.

With agents, real-time analysis of social media data across platforms like Bilibili and Xiaohongshu can be achieved in just 10 minutes, generating marketing trend insights and accurately capturing popular content styles. This is a stark contrast to traditional tools that may require at least a week. Additionally, agents can automatically generate compliant copywriting, reducing manual creativity costs.

03 Challenges and Misconceptions in Large Model Applications

Focusing on niche scenarios to form a closed loop, empowering the entire value chain with AI, and even generating agents may seem straightforward, but before these enterprises found solutions, they inevitably encountered detours. To this end, also discussed specific methodologies with enterprises for industry insiders' reference.

First, avoid blind worship of technology.

While large model advancements are crucial, Huang Yan emphasized that many enterprises only integrate AI into business scenarios when basic large models are highly advanced. In reality, current large model computational power is sufficient. Vertical enterprises need to transition from lab technology research to specific application scenarios, breaking down software and hardware barriers to achieve a business closed loop. "Don't be afraid of hard work, don't isolate yourself, and practice and make mistakes often," summarized Huang Yan.

Additionally, many vertical enterprises emphasize which renowned large model they use and the amount of parameter support they have when promoting their products. In reality, niche enterprises should prioritize optimizing user experience, particularly reducing user costs.

Weidu AI specializes in creating research-oriented agents. Its founder and CEO, Yang Yuliang, stated that the core value of agents lies in context. To this end, they use agents with superior interaction experiences to provide users with precise and reliable academic and knowledge research services. For instance, students, researchers, and professionals often face difficulties in time-consuming material searches and knowledge analysis in unfamiliar fields when writing papers, reports, or collecting cross-field information. Now, using Weidu AI's self-developed Dimension X1: Intelligent Research Assistant, key information can be extracted quickly to obtain reliable research results.

Kunlun Wanwei shares a similar perspective. While the underlying model is crucial as the foundation for deriving powerful upper-level products, accurately grasping product-market fit and deeply understanding market demand are essential to create products that truly address user pain points. Moreover, enterprises must form a product value closed loop of rapid feedback, action, and verification in response to new market dynamics.

Secondly, data accumulation and privacy protection are crucial aspects.

Numerous interviewees emphasized that the training of AI models hinges on an extensive amount of high-quality data. The process of collecting, organizing, and annotating this data demands significant resources and time. Given its paramount importance, data security and privacy protection emerge as pivotal concerns. Consequently, it is imperative for enterprises to establish robust data management and security mechanisms to preclude data breaches and malicious usage.

Li Weihan underscores that the deployment of large models marks merely the inception of enterprises' AI transformation, particularly since the evolution of agents is a gradual process that cannot be achieved overnight. Therefore, enterprises must prioritize the long-term accumulation of data to facilitate continuous business evolution and fortify their competitive position.

Thirdly, AI cannot fully supplant creators.

Amidst the surge of AIGC, the debate regarding AI replacing creators persists within the industry. However, several interviewees shared the perspective that AI cannot replace human creators. Ma Zihan pointed out that the essence of content creation lies in conveying emotions, attitudes, and personal styles, which AI struggles to replicate. While AI may occasionally exhibit whimsical creativity, it remains a distinct form of expression. Content expression inherently thrives on diversity, with AI having its unique style and different creators possessing varied styles, appealing to distinct audience groups. AI can enhance production efficiency, but human creative expression retains its unparalleled value.

Fourthly, organizational adaptation presents challenges.

As leading enterprises augment their investments in AI technology transformation, the adaptation of organizational structures becomes equally imperative. An organization that genuinely embraces, comprehends, and adeptly applies AI transcends mere technology adoption. This necessitates the development of a novel talent cultivation system and organizational form. For instance, AI has become the paramount evaluation criterion at BlueFocus, influencing recruitment, promotion, and incentives. Presently, BlueFocus boasts hundreds of product technology experts and AI seed talents.

Lastly, China's AI applications are still in their nascent stages.

Regarding foundational large models, China trails the United States by merely 1 to 3 months. However, in the application of these models, China remains in its infancy. An Xiaopeng noted that by 2024, the United States will have approximately 28 AI-related unicorn companies, with 6 to 8 in the code generation sector. In contrast, while numerous Chinese companies leverage AI to empower their businesses, the revenue they generate or the scale of AI-native enterprises pale in comparison to the US market.

Nonetheless, from a long-term perspective, this does not predestine Chinese enterprises to lag behind the United States. Kunlun Academy informed that while American enterprises hold advantages in foundational large model technology and algorithm accumulation, China boasts immense potential in application deployment and data scenarios. Breaking through technological barriers and integrating with industry resources represent the path for Chinese enterprises to overcome the current impasse.

Undoubtedly, as an infrastructural component, the vigorous development of large models and their pervasive integration into various industries constitute a general trend. An Xiaopeng anticipates that large models will drive all intelligent hardware, reconstruct all software, and activate all data. Despite China's nascent AI application ecosystem, amidst the open-source wave, small and medium-sized enterprises are achieving differentiated breakthroughs through "scenario-focused cultivation + data accumulation." As An Xiaopeng aptly stated, competition in the era of large models is not a "100-meter sprint" but a "marathon endurance race," offering Chinese enterprises a strategic window to cultivate a local ecosystem.

*The lead image was generated by AI.