General-Purpose or Vertical? SMB AI Firms Face an Existential Crossroads

Large models, led by ChatGPT, have rekindled the AI boom.

The large-model market bifurcates into general-purpose and vertical categories, distinguished by audience, use case, and application scope.

General-purpose models target foundational-layer breakthroughs. Baidu's Ernie Bot, Alibaba's Tongyi Qianwen, and iFlytek's Spark Model—all benchmarked against ChatGPT—belong here. Vertical models, by contrast, address domain-specific problems through product development, training specialized models atop general ones for finance, healthcare, education, elderly care, and transportation.

The virtue of general-purpose models is breadth: they span an expansive user base and a vast array of scenarios.

For specific use cases, however, enterprises require not omnicompetence but precision and quality.

Vertical models exploit this gap. Leveraging proprietary industry knowledge and partnerships with general-model providers, they train domain-specific models.

"Industry clients prize customization and engineering-delivery capability above all," said Wu Bingkun, CEO of Zhongshu Xinke, in an interview with Leiphone.

A vertical-model player, Zhongshu Xinke was founded in early 2021 by CloudWalk Technology, Xiamen Torch Venture Capital, and Minsheng E-commerce.

Zhongshu Xinke brands itself as a 'knowledge intelligence' firm within AIGC, engineering accumulated industry data and expert knowledge from the digital-city sector.

In essence, Zhongshu Xinke does one thing: fine-tune large models into domain-specific ones that drive industry efficiency.

Wu believes vertical industry models represent a significant commercial opportunity.

High barriers in general-purpose models steer startups toward vertical opportunities

Commercialization has long plagued the AI industry. Large models now offer a fresh path forward.

Wu likens AI to an oil rig: "No rig, no black blood for the industrial age. In the data age, no AI means untapped data value. Large models unlock infinite possibility."

Two years ago, digital-city AI relied on narrow models—facial recognition or license-plate recognition only. Products were constrained, R&D costly, and scalability limited.

With the breakthrough of large-model AI, data and AI are far more tightly coupled, and prior constraints have been dismantled.

That large models will disruptively reshape every industry is now consensus. Baidu's Robin Li, Alibaba's Daniel Zhang, and others echo the same refrain: In the age of AI large models, every industry application is worth rebuilding.

Judging by the wave of domestic large-model launches, foundational-model innovation remains a big-company arena.

Alibaba's Zhang noted that trillion-parameter model development is a全方位 'AI plus cloud computing' contest spanning algorithms, compute, networking, big data, and machine learning—a complex systemic undertaking.

General-purpose models demand vast data and compute, requiring massive AI infrastructure. Extended training and inference, in turn, entail hefty costs.

Hence, today's large-model market is dominated by mobile-internet titans—Baidu, Alibaba, Tencent—each with its own offering.

Big players target mass audiences, uniformly investing in NLP, CV, and cross-modal models, pre-training on similarly vast datasets.

This 'copycat' dynamic signals prowess but breeds homogeneity, eroding differentiation and competitive edge.

The 'all-encompassing' allure of large models notwithstanding, technology alone is insufficient. Value emerges only when technology becomes a product that delivers measurable user experience and conversion.

Put differently, 'large' and 'general-purpose' may sound attractive, but they fall short for enterprise clients.

Enterprise clients adopt large models to elevate their business. What they need is not technological dominance but products that excel at specific functions and maximize utility.

Clients pay for features they can productively use, not for those they cannot.

Small and medium startups see their opening.

Capital- and technology-constrained, they cannot enter the general-model arena. But proprietary industry knowledge gives them a natural edge in vertical exploration.

In the AI wave, small companies thrive by targeting niche tracks atop general-model capabilities, centered on the pragmatic question: 'how to use large models well.'

The Triple Crucible for Specialized Models: Technology, Scenarios, Data

Most mainstream AI large models do not expose training or fine-tuning. The prevailing approach is to offer an API after model development.

"Many AI majors are not yet opening up fine-tuning capabilities for industry-specific models," Wu observed.

First, large-model development is a progression from compute to framework to model to application. AI majors are still fortifying foundational capabilities.

Second, industry deployment requires integration with external applications—an area where domestic models remain immature. For now, AI majors lack robust engineering-delivery capability.

Against this reality, vertical-model players face a steep challenge fine-tuning and inferencing atop general models for customized solutions.

Through its ties to CloudWalk, Zhongshu Xinke participates in CloudWalk's model development and can train and fine-tune proprietary models atop CloudWalk's foundation.

"CloudWalk gives us easier API access, a more open partnership structure, and tailored integration," said Wu. "Apart from the AI majors, no other startup has this—it's our edge."

Wu identifies three factors critical to large-model deployment: AI R&D capability, accessible industry scenarios, and high-quality domain data.

On technology and scenarios, Zhongshu Xinke's three shareholders offer complementary strengths, creating a cohesive闭环.

CloudWalk backs algorithms and compute; Xiamen Torch Venture Capital provides a manufacturing digitalization sandbox; Minsheng E-Commerce supplies financial and commercialization support.

Of the three, industry data acquisition is the hardest—it directly dictates iteration speed and competitive positioning.

Industry data divides into static and dynamic categories by variability.

Static data—stable, non-real-time, with clear acquisition channels—includes internal documents held by governments, SOEs, and enterprises, plus database records.

Dynamic data—generated in real time across industry scenarios—constantly updates and shifts. Hard to obtain, it is the key differentiator from competitors.

Real-time dynamic-data access is a core advantage for Zhongshu Xinke.

Over the past two years, Zhongshu Xinke's 'iCity Life Service Platform' has reached 30 million individual users and hundreds of thousands of enterprise users across six provinces and 16 cities, amassing vast data.

Through the platform, Zhongshu Xinke has gathered rich scenario data, G/B/C user needs, and industry know-how, enabling it to deliver knowledge-intelligence engineering products and KAAS services atop mainstream large-model technology.

Moreover, general LLMs face a scarcity of domain-specific training corpora, resulting in knowledge limitations, cognitive biases, and memory hallucinations during niche deployment.

Knowledge limitations and cognitive biases can be mitigated through data accumulation. The harder challenge is memory hallucination.

The root cause: language models are not databases with true recall. They learn text-sequence distributions from training data and generate content accordingly, without retaining past inputs.

"Large models are not built overnight—they evolve," said Wu. "Zhongshu Xinke mitigates hallucination through context-aware methods and ongoing RLHF with clients."

Commercializing Specialized Models: A B-to-C Race

Unlike general-purpose models, industry-specific models demand faster commercialization as a safety net.

"In the industry-landing race for AI large models, speed is opportunity," Wu concluded.

Zhongshu Xinke's playbook: G-end for platforms, B-end for experience, C-end for rapid replication.

Entering via the G-end enables rapid market coverage and bulk aggregation of B- and C-end resources, channeling B-end experience into C-end replication.

"Only the C-end delivers viral replication—it is our priority now and in the next phase."

In urban life services, Zhongshu Xinke targets three deployment scenarios: education, elderly care, and cultural tourism.

In education, Zhongshu Xinke built a domain-specific model atop CloudWalk's Congrong model for schools and training centers, now piloted in parts of Xiamen.

Deployment follows three steps:

Step one: accumulation and annotation. Domain-specific teaching corpora built from years of digital-platform operations are annotated by experts. Step two: training and fine-tuning. Using CloudWalk's Congrong model, techniques such as knowledge distillation, weight quantization, and pruning transform a general teacher network into an industry-specific student network. Step three: deployment and feedback. User feedback is gathered in operation, and RLHF iteratively sharpens the model.

Of the three steps, expert annotation and RLHF are the two unavoidable hurdles. Overcoming them demands broader industry application, deeper knowledge accumulation, and relentless iteration.

Aligned with progressive model specialization, Zhongshu Xinke pursues an incremental trajectory from 'digital assistant' to 'digital avatar' for teachers.

For now, the model serves as a 'digital assistant'—auto-generating lesson plans, exams, and personalized student assessments. Teachers merely review and approve.

The 'digital assistant' frees teachers' time and boosts efficiency. Through co-work, it learns and evolves into a 'digital avatar,' approaching the caliber of an exceptional teacher.

China's education supply is strained, with a wide gap between advanced and lagging regions. An education model can transmit expertise from leading schools in developed areas to resource-poor regions via AI.

Zhongshu Xinke pursues a dual-track strategy: 'digital assistant' in advanced regions, 'digital avatar' in lagging ones.

"Knowledge from advanced-region teachers may underwhelm in equally advanced settings, but it largely suffices in underserved areas."

In short, a 'digital assistant' from developed regions functions as a 'digital avatar' in resource-scarce ones.

Wu disclosed to Leiphone that Zhongshu Xinke will deploy 'digital avatars' in Heilongjiang's underserved areas in H2 this year.

Going forward, as the education model matures, personalized instruction becomes a key evolution path. The 'digital avatar' can enter homes, tailoring education to each student and easing the family-education burden.

Conclusion

For now, neither general nor vertical players have achieved market dominance—all are sprinting.

Wu predicts: "A breakout AI model product will likely produce a winner-takes-all outcome."

For startups like Zhongshu Xinke, two pressures are clear:

First, rapid industry evolution demands blockbuster products and fast iteration—raising the bar for talent, product strategy, decision-making, and capital reserves.

Second, diverse entrants are reshaping the competitive landscape—in education, iFlytek and Yuanfudao have already joined the fray.

The industry-landing race for AI large models is, by nature, a sprint against the clock.

Original content by Leiphone. Unauthorized reproduction prohibited. Refer to republication terms.

Saved by 1 reader

Share: