Dahua Technology, a pragmatic AI player, is sprinting towards large-scale implementation of large models.

The sudden popularity of DeepSeek has led many to believe that the widespread adoption of large models has truly reached a turning point.

Behind this high - profile evaluation lies a set of impressive enterprise user data. According to statistics from Ai Analysis, as of February 21, 2025, 45% of central state - owned enterprises had completed the deployment of the DeepSeek model. At the same time, many industrial software manufacturers also connected to DeepSeek's API.

However, do these impressive data represent the truth of industrial implementation?

"Connecting to the large - model API" does not equal true "industrial implementation". The business logics of different industries and scenarios vary greatly. Even a powerful general large model like DeepSeek, without intensive training in industry - specific knowledge, cannot intuitively and quickly meet the needs of different industries as expected.

Since large models swept through industries two years ago, more and more people have realized that it may not be the top - notch AI research institutes and laboratories that can help enterprises complete the last - mile implementation, but the AIoT manufacturers that have accompanied industries for decades.

In the industrial circle, the name "Dahua Technology" is not unfamiliar. This "veteran" in the AIoT industry has applied more than 10,000 intelligent products centered around vision technology and over 500 solutions to more than 8,000 niche scenarios.

In the era of large models, how to make the high - brow "Transformer", "MoE", and "Agent" step out of laboratories and academic papers and truly take root in the industrial field has been a consistent proposition for Dahua.

On March 10, 2025, Dahua Technology released the Xinghan Large Model 2.0. In the words of Yin Jun, the dean of Dahua's Advanced Technology Research Institute: AI has truly become a "doer".

1. In the industry, AI is not just a set of algorithms but a system

On October 24, 2023, when the Xinghan Large Model 1.0 was just released, the general large models were at the height of their popularity, yet Dahua chose a professional and vertical path. The emergence of general large models like ChatGPT quickly raised the industry's expectations for them. Yin Jun noticed that in the past two years, many large - model products on the market were based on the concept of "generality".

At that time, vertical models focusing on specific industries were not an attractive story in the face of the grand pursuit of AGI and the technological race of piling up parameters. Many people also questioned whether it was appropriate to implement projects without using general models.

Before drawing a conclusion, let's first look at the truth of industrial implementation.

After years in the industry, Yin Jun is well aware of the difficulties in B - end business: B - end business is not only complex but also highly fragmented in processes and scenarios. Take the energy industry, which has a high incidence of safety accidents, as an example. It is an established fact that large models can play a role in production - line inspection and safety early - warning. However, safety management involves multiple links such as monitoring and analysis and requires accurate identification and understanding of complex production environments.

This means that a model trained with a single algorithm can hardly cover all processes end - to - end and has insufficient understanding of real - world scenarios. This weakness is also evident in the practical application of general large models.

Another key factor hindering the implementation of general large models is ROI (Return on Investment). Generality often means bulkiness and high cost. "The acceptable cost for customers on the market is between 50,000 and 2 million yuan," an industry insider once told 36Kr. However, the deployment cost of general large models, which can easily reach millions of yuan, often exceeds users' psychological thresholds.

All these lead to a conclusion: General large models are not the optimal solution for industrial implementation.

Since entering the large - model track, Dahua, which understands the path of industrial implementation well, has taken an unconventional approach and determined the route of industry - specific models. "When we started developing large models, we were clear about one thing: For large models to be industrially implemented in the future, they must be business - driven and tailored to vertical industries," Yin Jun told 36Kr. "So, on the one hand, we made the large models smaller to make the model computing power controllable; on the other hand, we continuously reduced the overall computing power consumption of applications to rapidly expand the scale of implementation."

This understanding also determines that the upgraded Xinghan Large Model 2.0 in March 2025 is not just a set of algorithms but a system.

For example, the professional knowledge and scenario data of many enterprises are scattered in their previous systems and traditional AI models. The solution centered around the Xinghan Large Model 2.0 includes both the collaboration between the large model and the enterprise's traditional systems and the cooperation among niche - scenario models of different scales. Dahua's practical experience shows that only by integrating technology into a business system can AI truly operate in business scenarios from an abstract technical concept.

In the communication with 36Kr, Yin Jun mentioned "meeting users' needs" in almost every sentence. This also represents Dahua's attitude from the construction of the model system to the productization process - instead of "looking for nails with a hammer" and making customers adapt to technology, technology should address the pain points of the industry.

A typical example is that the birth of the three series of the Xinghan Large Model 2.0 is all related to customers' needs:

The V - series visual large model, which excels in multi - modal capabilities, is applied to tasks such as ultra - small target detection and complex - scenario recognition in high - frequency industries like urban governance and manufacturing. The M - series multi - modal large model, as the "brain center", stems from the needs of many customers for text - to - image and image - to - image searches in actual business. The L - series language large model, which plays the role of a "conductor" in actual business, is developed to meet the needs of many customers for a human - machine interaction model based on language and text, according to Yin Jun.

The research and development of the large - model series address the issue of the "superstructure" of technological infrastructure. When it comes to productization, it is not easy to effectively integrate the model capabilities into the enterprise's workflow.

In the long - term process of dealing with customers' needs, Yin Jun found that in the workflows of many enterprises, the subsystems at different levels are often intertwined, and these subsystems are connected to the APIs of different tools and databases. The complexity and fragmentation of the systems make it difficult for large models to fully utilize the knowledge and tools scattered in them during the implementation process.

Therefore, "the essence of large - model implementation is to connect these complex systems, tools, and data interfaces according to the enterprise's business," Yin Jun told 36Kr.

The industry intelligent agent and workflow engine developed by Dahua can be summarized as dividing the model - implementation process into three steps: "disassembly, invocation, and orchestration". The existing complex business processes of enterprises are disassembled into an atomized operator library that can be orchestrated. According to the specific business and scenario requirements, the intelligent agent can quickly invoke the workflow engine. Finally, through the dynamic combination of algorithms and workflows by the engine, the efficient adaptation of technological capabilities to business needs is achieved.

This set of "atomized" disassembly and orchestration workflow solutions has been applied to various industries, quietly changing people's lives. Take the urban emergency command and dispatch scenario as an example. Once a fire breaks out, the intelligent agent can quickly access the surrounding surveillance cameras, assign rescue tasks according to the individual - soldier equipment, and initiate audio - video consultations through the integrated communication system to activate the emergency response plan.

In the future, there will be more industry stories about the transformation from "experience - driven" centered on humans to "cognitive - intelligence - driven" centered on AI.

2. Dahua has provided an answer to industrial implementation with over 30 years of experience

Since its establishment over 30 years ago, Dahua has been deeply involved in technological fields such as IoT perception and computer vision, aiming to solve the pain points related to "seeing" in the business process.

An interesting echo is that in the technological wave, Dahua was among the first to "see" the needs of enterprise customers.

"The ability to see" is a high - frequency, basic, yet often overlooked need by technology providers. For example, in the power - inspection scenario, many substations are located in remote mountains, making manual inspection extremely difficult. To achieve unmanned operation, the first problem to solve is to "accurately read" the electricity meters.

The industry's strong demand for "seeing" has become the most natural reason for Dahua to enter the large - model era. Yin Jun told 36Kr that Dahua started with visual understanding. "More than 90% of Dahua's technologies, including deep - learning models, are centered around video."

Discovering customer needs is the starting point of business, while technological strength is the hard power to stand in the industry. A surprising fact is that in 2019, nearly four years before ChatGPT triggered a storm, Dahua Technology was sensitive enough to introduce Transformer into the company and used ViT (a vision model based on the Transformer architecture) technology to develop a semi - automated annotation solution. By early 2020, Dahua had successfully developed an automated annotation model.

In the large - model track, those who have entered the field deeply understand that technological research and development not only requires continuous investment but also depends on a strong team.

Let's first look at the investment. Over the years, Dahua's R & D investment has remained above 10% and shows an increasing trend year by year. According to the financial report, in 2024, Dahua's R & D investment reached 4.21 billion yuan, accounting for 13.09% of the total revenue. Yin Jun told 36Kr that over the past decade, Dahua's algorithm team and computing - power scale have been continuously expanding along with the scope of business.

Now, let's look at the team. As the team behind DeepSeek has been exposed to the public, young talents have gradually become the focus of enterprise recruitment. Since Dahua has focused on AI technology for more than a decade, the cultivation of young talents has become the core mode of team building.

According to Yin Jun, Dahua's algorithm team recruits fresh graduates every year, and the proportion of master's and doctoral students is as high as 98%. "We will gradually train fresh graduates and let them grow into our core backbones."

It is worth mentioning that different from the concept of "being isolated from the outside world" and focusing solely on scientific research, Dahua not only requires young talents to understand algorithms but also encourages them to step out of the laboratory and experience the actual implementation process of business. For example, every fresh graduate who joins the algorithm team is required to complete the on - site implementation and delivery of a project during the internship. "Only when they have seen and touched their products on - site and heard users' opinions, even complaints," Yin Jun said, "can they truly understand how to improve the technology and products they are working on."

Generations of talents have continuously moved towards the field of business implementation, ultimately forming Dahua's solid business accumulation and industry understanding over the past 30 years. In 2023, when the large - model craze was surging, Dahua's "1 + 2" artificial - intelligence capability system - a set of architectural systems plus scientific research and engineering capabilities - had embedded AI technology into more than 8,000 niche scenarios.

"We understand business, have experience, and know how to serve customers and make them better operate this system," Yin Jun summarized the essence of technological implementation. "Whether it is digitalization or intelligence, there is a digital system behind it. How to make the digital system better meet customers' needs and play its role in customers' business processes is what we are good at."

In the large - model track, there are many participants with different strategies. Some are striving to reach the peak of AGI, some are riding the wave at the forefront, and some are pursuing practical implementation. The wave is both an opportunity, and on the contrary, a little slack will lead to being washed out.

From the digital era to the intelligent era, with solid technological strength and a wealth of business experience, Dahua has carved out a unique position in the large - model field.

3. If large models are well - implemented industrially, they are not a bubble

In a mine shaft that is thousands of meters deep, how can we accurately identify potential safety hazards in vehicles, equipment, and coal piles in the complex underground working environment? Facing dozens of high - frequency accidents every year, how can we manage the work norms of dozens of workers to prevent accidents?

Before the digital era, these series of problems had always been obvious pain points in the coal - mining industry. Yin Jun remembers that initially, vision technology could only solve the problem of personnel management, helping enterprises confirm whether workers were wearing safety helmets and other protective equipment and whether they were working in safe areas.

Later, with the development of large - model technology, the scope of what Dahua can help enterprises "see" has extended from personnel management to the management of various processes such as vehicles, conveyor belts, and coal piles, almost covering the entire process from pre - control to underground operations and logistics transportation.

The sense of value of technology - based enterprises not only comes from the improvement of technological capabilities but also from the expanded implementation space after the improvement of their own technological capabilities. Regarding the implementation of large models, Yin Jun has a practical "greed": "After entering an industry, as you gain a deeper understanding of the business, you will want to help customers do more and improve the entire business process end - to - end. With the development of technology, we can use new technologies to upgrade the industry again, which is also our direction of effort."

Covering an entire industry from individual scenarios is Dahua's ambition in the large - model era. To accelerate the achievement of this goal, Dahua has its own considerations in choosing the main scenarios to enter.

On the one hand, it is to find "difficult but correct" scenarios. In Yin Jun's view, these complex business scenarios often represent strong enterprise needs that technology has failed to solve for years. Once the technological model can be successfully implemented, it can be replicated in the industry, generating multiplied value.

The smart move Dahua made was to cooperate with leading enterprises in the mining industry. Yin Jun told 36Kr that at the beginning of entering the coal - mining industry, Dahua had in - depth discussions with leading domestic coal - mining enterprises about the possibility of cooperation. "We hope to jointly create some industry benchmarks or things recognized by the industry." Now, the implementation effect of this model solution in the coal - mining industry is obvious. For example, in core functions such as large - object detection on conveyor belts and deviation monitoring, the accuracy rate has increased from 80% to 93%, and in harsh environments such as high - temperature and poor - lighting conditions underground