The realm of technology and innovation is constantly evolving, but at its core remains a singular question: How can these advancements be transformed into tangible commercial value? As massive models become more prevalent in practical applications, the industry has witnessed a significant surge in the launch of specialized large modelsBy December 2024, Baichuan Intelligent will unveil its comprehensive financial large model, Baichuan4-FinanceIn parallel, Zhongguancun Science and Finance has introduced its upgraded model platform 2.0, while previously, Qifu Technology disclosed its efforts to implement large models in microfinance scenarios.

In stark contrast to the fierce competition among tech giants in the realm of generalized large models, these specialized industry-focused models emerge as far more relatable and applicable.

"In 2024, numerous cloud vendors approached us, recognizing that we are at the front lines with practical business scenarios

This allows us to build large model products that can receive immediate feedback from users," shared an executive from a fintech company during an interviewSuch statements underline the increasing collaboration between technology companies and traditional industries.

Recently, Zheng Weimin, a member of the Chinese Academy of Engineering and a professor in the Department of Computer Science and Technology at Tsinghua University, commented in an interview that the demand for generalized large models is not as high"I believe we only need three or four comprehensive models across the nationThe next step in the development of large models is to enhance their applicability and foster a better software ecosystem,” he stated.

Yet, the challenges posed by substantial technological investments remainThe critical trifecta of computational power, algorithms, and data form the backbone of large model development, with both power and data recognized as significant barriers among enterprises today

• Pitfalls of the Forex Market
• Competition in the Billion-Dollar Data Market
• Manufacturing Sector Sees Three Months of Growth
• Industry Giants Push for Real-World Tech Deployment
• 2025 Treasury Market: Trends and Drivers

The growing desire for industry-specific large models arises from the hefty investments required in these three areas.

As the scale of large language models expands, with parameters reaching into the trillions, the financial, human, and energy resources needed for training sessions become increasingly prohibitiveZheng illustrated this concern during the interview, explaining that foundational large models necessitate considerable computational loads, extensive GPU resources, and substantial storage, whereas industry-specific models do not require such vast amounts of power.

He cited a well-known startup in the large model arena that focuses on maximizing data input and extending context windows for superior outputHowever, the high inference loads mandated the acquisition of more inference cards, leading to overloaded memory; in situations where usage spikes unexpectedly, the risk of system failures heightens

The executive disclosed that this startup had attempted to purchase additional computing power multiple times without success in resolving these challenges.

The stark reality remains that most organizations cannot sustain such costsIndustry insiders have elaborated that the large model ecosystem is divided into training and inference processes, where initial training incurs high expensesIn 2023, inference costs remained steep, but it was expected that by 2024, they would see a significant reduction due to technological advancements.

The data challenge emerges as a significant hurdle as wellPredictions from Epoch AI Research warn that by 2026, high-quality linguistic data currently utilized for training AI models will be depletedResearch from information technology analysis firm Gartner further emphasizes that by 2030, synthetic data will become the primary source of training data for AI models.

There is widespread agreement that the development of large models is transitioning from a fervent growth phase into a more meticulous implementation stage.

Zheng Weimin points out two distinct characteristics that define the era of large models in AI: the evolution from unimodal to multimodal foundational models and the acceleration of intelligent upgrading across industries facilitated by large models

He envisions three categories of enterprises poised to thrive in this new environment: those developing large models, those implementing them in practical settings, and those providing computational support for these models.

Industry-focused large model companies fall into the second category, and this focus is essential for fostering genuine advancements in commercial applications.

Yu Youping, president of Zhongguancun Science and Finance, echoed this sentiment in a recent interviewHe asserted that the market primarily seeks practical tools that return to the essence of businessFor industry-specific large models to succeed, they must deliver end-to-end solutions that address real problems while also providing a full range of services.

The financial sector stands out as a crucial domain for the deployment of large modelsIn November 2024, there was a clear push in the financial industry to enhance digital technology capabilities to facilitate the digital transformation of the sector

Data from the Hang Seng Research Institute indicated that in the first three quarters of 2024, the total amount of disclosed large model projects in China reached 2.075 billion yuan, an increase of 163% compared to the total for 2023. Notably, 66 projects in the financial sector brought in 1 billion yuan, making up 4.9% of the overall marketBy the end of November 2024, the number of financial sector projects had climbed to 103, with total funding increasing to 2 billion yuan.

While the financial sector's absolute proportion in this industry surge may not seem substantial, its higher standards for technology and security mean that the growing capabilities of large models within finance often set the groundwork for similar applications in other fields, thereby creating a “backward-compatible” dynamic.

The aforementioned fintech executive remarked on how their management team clearly identified the need to discover the most efficient applications of large models within fintech during the 2023 OKR setting process.

There is a strong consensus on the importance of recognizing the value derived from use cases

Fei Haojun, chief algorithm scientist at Qifu Technology, emphasizes that if technology accounts for 40% of the large model's effectiveness, the understanding and cultivation of relevant business scenarios must represent 60%. He identifies three essential components for successful large model products: application scenarios, a data flywheel, and intelligent agents.

Yu Youping articulated that a combination of “platform + application + service” establishes the most effective path for large model deploymentHe pointed out that the differences between platforms hinge on three critical aspects: the magnitude of computational power, the speed of model response, and the capacity to build applications represented by intelligent agentsWithin the framework of intelligent agent application development, scenario templates emerge as vital elements.

How can we comprehend the significance of these scenarios? Yu explained that while the methodologies for understanding scenario value and knowledge abilities may be universal, the data and industry contexts are not interchangable

For instance, the sales processes for insurance and wealth management products are inherently different, leading to the essential conclusion that even within marketing calls, divergent scenarios arise from varying industry demands.

Under certain conditions, these scenario capabilities might allow for portability, particularly in customer acquisition marketingYu offered an example where extracting vertical scenario data and knowledge from one industry to infuse into a relevant industry model could enable post-training use, thereby achieving standardizationThis potential for capability migration and standardization opens up vast possibilities for the application of large models across different sectors.

Currently, in the financial sector, large models primarily function to empower internal operations and enhance customer service, outbound calling, and customer acquisitionYu reports impressive results in an intelligent marketing scenario, revealing a conversion rate of 3.5% for outbound calls in promotional campaigns, representing a substantial 130% improvement over traditional AI call methods, and closing the gap to only 17% when compared to human call agents