The early days of WENGE AI (HKEX: 01956) were not marked by the clear market positioning it holds today.

In the winter of 2018, the team was stranded on a highway late at night due to a heavy snowstorm during a business trip to northern China. Such arduous travel was not uncommon back then; to make clients truly understand what problems artificial intelligence could solve, the team had to visit them one by one, explaining its potential industry by industry.

At that time, mentioning AI or algorithms typically elicited reactions like, "This doesn't seem very relevant to me."

Seven years later, on June 26, 2026, WENGE AI listed on the Hong Kong Stock Exchange.

In the grey market session on June 25, WENGE AI closed at HK$115.4, representing a 90.12% increase from its IPO price of HK$60.7. Its Hong Kong public offering was oversubscribed by 5,966.78 times, and its international placement was oversubscribed more than 20 times.

On its first trading day, the company's stock surged by as much as 87% intraday. In 2025, WENGE AI's revenue surpassed 400 million yuan, with a net dollar retention rate from existing customers reaching a high of 139.5%.

By revenue, the company ranked first that year among China's enterprise-level decision intelligence service providers powered by large models, capturing a 10.2% market share.

Unlike the early days when explaining the utility of AI was a constant struggle, the company's frontline sales staff recently noted a distinct, proactive warming of demand during intensive visits to dozens of small and medium-sized enterprises in the Jiangsu-Zhejiang region.

Clients now express immediate purchase intent upon meeting. For instance, in a textile factory, the product serves not only as an AI-powered diagnostic assistant for faults but also as a decision-support tool for management. A foreign trade business owner can simply ask a question in natural language, and the system automatically links underlying data to generate a decision-making report.

This shift is underpinned by an eight-year development effort that produced a technical framework named DOMA. This framework organizes data, industry knowledge ontology, large models, and intelligent agents into a complete pipeline, enabling AI to be embedded into every production and decision-making step within an enterprise.

In the company's assessment of the next stage of AI capability, the truly important question is no longer just "Can AI answer questions?" but rather "Can AI further simulate the world and predict the future?"

Based on this judgment, WENGE AI recently launched its general-purpose decision-making large model, Decitron.

The company's co-founder and CEO, Luo Yin, stated that the core of Decitron is not about answering questions but about transforming the complex world into a system that is computable, simulatable, and verifiable.

Here is the complete story of this "Chinese Academy of Sciences-affiliated AI company" from laboratory to IPO.

The Initial Challenge: Finding the Right Problem to Solve

The story begins in Zhongguancun.

Before 2017, the core founding team members were researchers at the Institute of Automation of the Chinese Academy of Sciences. It was during this period that the team gradually formed a clear consensus: the value of artificial intelligence should not be confined to laboratories and academic papers but should advance into industrial practice, translating into real business and economic value.

In 2016 and 2017, national policies were introduced to encourage researchers to leave their posts to start businesses. Against this backdrop, doctoral-level researchers like Wang Lei, Luo Yin, and Zeng Dajun co-founded WENGE AI.

Among them, Wang Lei serves as co-founder and chairman, Luo Yin as co-founder and CEO, and Zeng Dajun as co-founder.

Wang Lei recalled that the Academy told them, "If the venture doesn't succeed, you can come back." He added, "Later, we found we couldn't go back. The team had ventured out together, the company had grown, and with so many colleagues, everyone needed to work here and create their own income."

The company's name, "WENGE," is derived from a classical Chinese phrase meaning "understanding the refined meaning upon hearing the music." Its mission sounds ambitious: to build advanced business and decision systems for the AI era.

Despite its CAS background, at the outset, the company had no clients, no sales team, and virtually no one who understood what they were talking about.

"With our background, many clients did trust us. But when it comes to real market competition, I believe whether you come from CAS or Tsinghua, everyone competes on the same stage," Wang Lei said. "Ultimately, it's about providing valuable services to clients."

The company's first foothold was the media and communications industry, a sector not considered particularly attractive from a capital markets perspective at the time.

However, WENGE AI had its own rationale: in 2016 and 2017, today's general-purpose large models did not exist; technology was built on small-to-medium scale data and specialized machine learning. The media and communications field happened to possess vast amounts of complex, multimodal data. Text, images, video, and audio all required algorithms for recognition and understanding, and there was a clear willingness to pay.

The team's thinking then was that they first needed to find a "nail" they could hammer in, upon which they could build a larger business landscape.

Helping large media institutions with complex data analysis and AI-assisted decision-making became that first nail.

At the time, a major challenge for large media organizations was the overwhelming volume and variety of daily interview materials—video, audio, text in different formats—that were difficult to utilize and store effectively.

Addressing this pain point, WENGE AI performed semantic annotation, storage, and tagged retrieval on interview materials like images, audio, and video for later reprocessing and dissemination.

While this might sound mundane, establishing such fixed workflows indeed brought efficiency gains for large media groups handling massive daily volumes.

From there, the company step-by-step built full-chain service capabilities from data management and content analysis to topic decision-making and communication effect evaluation, gradually securing several media clients.

Although the industry's growth potential was limited, it became a lifeline for WENGE AI during difficult times.

In the toughest year of 2020, it was this client base that provided relatively stable cash flow, preventing the company from falling during the earliest, coldest winter of AI commercialization.

For WENGE AI, this experience yielded not just clients and revenue but also foundational capabilities later proven invaluable: the habit of handling multi-source heterogeneous data, systematic methods for modeling "event-object-relationship-impact," and engineering experience ensuring reliability in high-governance scenarios.

With the media sector as its base, the company then expanded into government affairs, finance, scientific research, and industrial fields.

Wang Lei used an analogy to explain the improved migration efficiency: "Previously, after building System A, you'd have to start from scratch for System B. Now, after A, moving to B might leverage 70% of the accumulated work, requiring only 30% new effort. After doing more, you find maybe 90% is common."

Looking back, the true importance of that initial "nail" was that it allowed the team to first build the capability to handle complex organizational data, understand business rules, and construct industry knowledge systems.

These capabilities were almost entirely reused when entering government, finance, research, and industrial sectors later on.

A Crucial Investment Decision

In late 2022, ChatGPT burst onto the scene. At the time, WENGE AI's core team held meetings until the early hours, discussing the potential impact. They concluded this was not merely a product-level innovation but possibly a shift in the underlying technological paradigm.

"The old model involved specialized machine learning solving specific industry problems; an algorithm working in domain A might require retraining a completely different model for domain B," they noted.

In Wang Lei's view, post-ChatGPT, large models changed the entire competitive logic of the AI industry.

The shift moved from "training one model for one problem" to "building a general intelligent foundation capable of understanding the world." Missing out on this round of foundational model capability building could mean losing eligibility to compete in the next stage.

The team then faced a major choice: should they train their own large model?

Training a model meant a multiplication of R&D investment, but failing to do so risked obsolescence.

"The company is about to become profitable, and now you're pouring money into this. Can you earn it back?" some investors questioned at the time.

After repeated deliberation, the team's judgment was: "If we don't do this, there might be no future for WENGE AI. In a few years, WENGE AI could be eliminated from this race."

In early 2023, Wang Lei and Luo Yin gathered fellow researchers from the Institute of Automation to discuss plans, ultimately deciding to go through the training process themselves.

The cost of this "big bet" was clearly visible on the financial statements.

That year, WENGE AI's R&D expenses soared to 180 million yuan, with a net loss reaching 260 million yuan for the same period.

Behind the widening loss, the company also filled in the crucial puzzle piece of foundational model capability.

In June of that year, WENGE AI launched its fully self-developed "Yayi" large model, becoming one of China's earliest enterprise AI companies with complete foundational model training capability.

Foundational model capability thus became a key variable in the company's market competitiveness.

According to the company, after 2023, during client project engagements and bidding processes, it gradually perceived a differentiation in enterprise AI competition: companies without foundational model accumulation often faced capability limits in designing technical solutions for complex scenarios; while companies with model R&D capability might not necessarily deeply understand industry-specific business and decision-making loops.

What enterprises truly need is not just model capability itself, but a systematic ability deeply integrated with business scenarios.

WENGE AI's current model system is divided into two categories: general-purpose large models like Yayi, and specialized large models for specific scenarios, like Panshi.

The underlying philosophy is "integration of general and specialized capabilities," where foundational models address common sense and general problems, while specialized models tackle deep needs in vertical domains.

It's difficult for a single vendor to dominate with one model. Those good at foundational models may not excel in scientific research or code models. The WENGE AI team believes the model ecosystem will ultimately form a hierarchical structure of foundational models, industry-level models, and professional models.

Navigating the Final Mile of AI Implementation

If measured solely by general-purpose large model parameter scale and computing power investment, WENGE AI might still find it hard to match major internet giants and other AI players.

However, enterprise AI competition is not just a race of foundational model capability. For enterprise AI companies like WENGE AI, the real key lies in: whether they can combine large model capabilities with industry knowledge, business data, and organizational processes to solve complex decision-making problems in real-world scenarios.

Many enterprise AI deployments fail not because the models aren't smart enough, but because the models don't understand the enterprise's own business rules.

WENGE AI proposed the DOMA framework, an acronym for Data, Ontology, Models, and Agents. While the framework sounds academic, it is a crucial methodology refined from years of industry practice.

Data, ontology, models, and agents together form the underlying foundation for enterprise AI operation, with the most critical layer being the industry ontology. It enables AI to not only understand language but also understand how an enterprise thinks.

Taking investment analysis as an example, if a model is asked to evaluate an investment target, one query might use one analytical framework, and another query might use a different methodology.

But if the model is pre-instructed that evaluating a company requires focusing on indicators like financial statements, core team, and technological advancement, and then asked to perform quantitative and qualitative analysis within this framework, the consistency and accuracy of results improve significantly.

This process of "telling the model how to view data" is ontology modeling.

The ontology layer's role is to structure the business objects, relationships between objects, operational rules, and constraints within a specific industry, forming a set of "industry thinking patterns" that AI can understand.

Over the past few years, the greatest value of large models has been helping people acquire information and generate content. However, the problems enterprises truly care about are often not "what is the answer?" but "what should we do next?"

A technician involved in Decitron's development noted that many complex decision problems are difficult because they often involve multiple entities, multiple variables, and continuously changing constraints, making them irreducible to one-off Q&A.

"Genuine decision problems must first be transformed into a structure that is analyzable, decomposable, and simulatable." In his view, this is also a key distinction between decision engines and traditional generative models: the latter excel at generating responses based on existing knowledge, while the former emphasizes continuous modeling of event relationships, path changes, and outcome branches.

This way, when models view data, they are no longer like blind men touching an elephant but have a clear analytical framework.

The DIP Decision Intelligence Platform recently launched by WENGE AI is also a productized manifestation of this methodology.

DIP addresses the three most common problems in enterprise AI implementation: fragmented data, complex business processes, and difficulty in driving action.

In enterprise scenarios, a single order is often linked to an entire business chain involving customers, products, inventory, contracts, etc. If AI can only read data without understanding the relationships between these data points, it can hardly participate in genuine business decision-making.

DIP's function is to transform data scattered across different systems into "business ontologies" that AI can understand, then allow models to perform analysis, judgment, and action based on business rules.

For example, if a business user wants to know "which Grade A healthcare industry clients are at risk of churn in the next 30 days," DIP automatically links information such as customer tier, order history, and after-sales work orders to form a complete customer operation view, ultimately providing reports listing at-risk clients and explaining the reasons.

This means WENGE AI does not intend to compete head-on with giants on parameter scale. Its real battle is for the "final mile" of enterprise AI implementation: enabling models to read business processes, understand rules, and truly enter enterprise decision-making workflows.

Decitron: Shifting AI from Answer Generation to Decision Simulation

If the past few years have shown the world AI's capabilities in content generation, knowledge Q&A, and efficiency enhancement, then in WENGE AI's view, the next step for enterprise AI is moving from "being able to answer questions" to "being able to assist in decision-making."

This is also the important context for the company's recent launch of the Decitron decision engine.

As a general-purpose decision intelligence model product designed for complex, open scenarios, Decitron is not limited to a single industry or task. It focuses on complex decision problems involving uncertainty, multi-path choices, and multi-party games, helping users understand situations, simulate trends, compare options, and form more valuable judgments.

Its application scenarios can cover financial markets, macroeconomics, international relations, public governance, corporate strategy, investment research, and industry analysis, representing a concentrated productization of WENGE AI's years of decision intelligence capabilities.

Taking U.S. interest rate change prediction as an example, interest rate decisions themselves are not determined by a single data point but are influenced by multiple factors including inflation trends, employment data, economic growth, market expectations, policy statements, and the international situation.

Faced with such highly uncertain complex problems, Decitron assists users by performing continuous analysis of multi-source information and multi-path simulation to understand possible directions of rate changes, key variables, and their potential impacts under different scenarios.

This highlights the value of decision intelligence distinct from traditional information retrieval and generative Q&A: it focuses more on assisting user judgment in uncertain environments.

In more everyday-life scenarios, Decitron can also be used for "university application planning." When college entrance exam results are released, application planning often becomes a crucial choice for students and parents. For this scenario, Decitron launched a "university application simulation" function to provide auxiliary reference for planning.

Unlike some application tools that primarily offer relatively certain reference results like "aspirational, stable, safe" based on existing data and rules, Decitron emphasizes simulating the opportunities, risks, and differences behind different choices, helping users make judgments based on an understanding of multiple possibilities and their own preferences.

From a capital markets perspective, the significance of Decitron also lies in further strengthening WENGE AI's differentiated label of "decision intelligence."

Currently, competition among AI companies is gradually shifting from model parameters and general capabilities towards the ability to enter real business scenarios, solve complex problems, and achieve sustainable delivery. Compared to solely pursuing general conversational ability, WENGE AI has chosen a path more oriented towards the enterprise level, more focused on scenarios, and more emphatic on decision-making value.

For the company, Decitron is both a product launch and, at the listing milestone, an important window to showcase its long-term technological roadmap to the outside world. As AI moves from generative applications to decision-making applications, the competitive focus of enterprise AI may also shift from "who answers better" to "who helps clients make better judgments."

Advancing Deeper into Industry

For WENGE AI, listing on the capital market is not the end of the entrepreneurial stage but marks entry into another, longer-term competition.

Behind the impressive growth curve lies an unavoidable financial reality.

From 2023 to 2025, the company's revenues were 250 million, 318 million, and 405 million yuan respectively, representing a compound annual growth rate of 27.4%. However, cumulative losses over the three years reached 583 million yuan.

The core reason for the current losses is R&D investment, with cumulative R&D expenses nearing 500 million yuan over the past three years.

Nevertheless, one set of data is particularly noteworthy. The average delivery cycle for WENGE AI's AI services shortened from 185 days in 2023 to 80 days in 2025.

The shortening delivery cycle indicates a significant increase in the reuse rate of underlying modules. This is strong evidence of the company's system migration capability and is expected to become an important competitive moat in the future.

However, the enterprise AI market is intensely competitive. "Decision intelligence" as a market concept still requires time for broader adoption, and the progress of platformization needs more financial indicators for ongoing validation.

Wang Lei is clearly aware of this. In discussions, he referenced Palantir—a company that also started with data governance and decision support for complex organizations and has recently garnered significant capital market attention.

Simultaneously, WENGE AI is also monitoring the progress of global advanced AI firms like Anthropic. For the company, such reference does not mean simple replication but rather, based on the realities of China's market environment, client structure, and industrial digitalization process, finding a suitable path for its own enterprise AI implementation.

Wang Lei expressed that if one day they can truly succeed in the endeavor of "decision-making large models," he hopes WENGE AI can "leave a mark of AI memory for the era."

The launch of Decitron is a significant step forward for this ideal. For WENGE AI, which has just entered the capital markets, "general-purpose decision-making" is no longer just a technical direction but is becoming the core product narrative for its future.