DEEPEXI TECH has officially launched an upgraded version of its Deepexi Enterprise Large Model and the Deepexi OS AI-level enterprise operating system. The company also jointly inaugurated an Embodied Intelligent Brain laboratory with Tianjin University.
In an interview, Zhao Jiehui, Founder, Chairman, Executive Director, and CEO of DEEPEXI TECH, along with Yang Lei, Co-founder, Executive Director, and President of the Product and Solutions team, stated that the company focuses on a deep understanding of enterprise business logic. They aim to establish trusted boundaries for corporate AI through an "ontology paradigm" and to redefine service value measurement using a "token economy."
Mr. Zhao used an analogy: "A digital employee moving from Company A to Company B cannot take specific data with them, but they can take the ability to understand that data." In his view, this transfer of capability is the core value of implementing enterprise large models.
Addressing the common challenges businesses face with AI adoption, Mr. Zhao compared the situation to the recent "lobster" phenomenon at the launch event. He explained that while many companies want AI employees that can perform tasks, they find that general-purpose models often fall short in terms of cost, security, and accuracy. He believes enterprises need not just a simple large model deployment, but a standardized and controllable way to cultivate "AI employees" tailored to their specific business logic.
When asked about the positioning differences between DEEPEXI TECH, general large models, and sensational products like Open Claw, Mr. Yang Lei analyzed that while consumer AI products have strong explosive growth potential, their technical paths have inherent limitations when faced with the accuracy, compliance, and security requirements of enterprise-level applications.
Mr. Yang stated that DEEPEXI TECH's core barrier lies in constructing an ontology paradigm. This is a knowledge association framework for specific enterprise business scenarios, abstracting the business logic, management rules, and knowledge structures behind data into a semantic network understandable by models.
Furthermore, Mr. Zhao clarified that an ontology does not refer to specific enterprise data, but rather to a closed business loop, such as the knowledge logic and business semantic network within scenarios like equipment failure maintenance or supply chain adjustments. It encompasses the relationships between data fields, documents, drawings, and business processes. Previously, this work relied on manual efforts by FDE engineers. Through serving over 300 leading clients, DEEPEXI TECH has accumulated the capability to model this process.
He emphasized that the essence of this technical approach is to enable models to understand enterprise business and execute encoding, not merely to conduct conversations.
On ensuring accuracy and trustworthiness in enterprise AI, Mr. Zhao used the healthcare industry as an example, noting that all responses and actions must be traceable and evidence-based. For instance, in building an AI service for a specific business scenario for a utility client, DEEPEXI TECH worked entirely within a provided 800T dataset. By governing these multimodal data and supplying them to the enterprise large model, they ensure all answers are traceable and trustworthy. He believes the core difference between enterprise and consumer AI is that it must operate within constraints: "It can admit the unknown, but it cannot fabricate answers."
Mr. Yang further revealed that DEEPEXI TECH incorporates constraint mechanisms in model training. He stated that building these "data fences" results from years of experience serving leading clients. The company manually associates vast amounts of drawings, work orders, and knowledge logic to form high-quality datasets covering enterprise knowledge and business semantic networks, which are then used for model training. In his view, the lack of such specific scenario training data on the internet is why major tech players find it difficult to cover this domain.
Regarding the industry-wide concern about computing power consumption, Mr. Yang proposed a "token economy" perspective. He believes that as AI penetrates core business operations, traditional service pricing models based on person-days will gradually be replaced by computing efficiency models based on token consumption.
Mr. Yang noted that in comparative tests, DEEPEXI TECH found significant differences in token consumption for the same task across different platforms. Some consumer-oriented products, due to disordered skill invocation, could consume up to 10 times the tokens of DEEPEXI TECH's Deepexi Enterprise Large Model. He analyzed that the core of enterprise scenarios is not simply competing on computing scale, but focusing on the business efficiency generated per unit of computing power.
On the commercialization path, Mr. Zhao clearly stated that DEEPEXI TECH will not engage in reselling computing power, but will focus on the closed loop between token consumption and business value. He mentioned that the company currently uses a license model based on usage duration in overseas markets, but the underlying support is a more efficient token consumption mechanism. Mr. Yang added that as AI evolves from an auxiliary tool to a factor of production, enterprise resource allocation will shift from human resources to computing resources. How to generate clear business returns from computing power consumption is a key question future enterprise service providers must answer.
The core product launched, Deepexi OS, is positioned as an "AI-level enterprise operating system." According to a live demonstration by Mr. Yang, the system integrates the complete chain from data ingestion and business comprehension to task execution. Through the enterprise data fusion platform FastData Foil, heterogeneous information like engineering drawings and operational status data is parsed and fed into the model. The Deepexi Enterprise Large Model then generates enterprise-specific business models based on the ontology paradigm. These models are subsequently converted into composable skill modules via the FastAGI Enterprise Agent Platform, ultimately orchestrated into AI employees for specific scenarios.
It is reported that the system currently designs 108 business ontologies and has released over 280 skills, covering five major sectors: manufacturing, retail, healthcare, transportation, and general use.
Notably, in a manufacturing operations demo, an AI employee autonomously completed a full closed-loop process from equipment failure alarm and root cause diagnosis to maintenance plan formulation and work order dispatch. Upon detecting abnormal equipment temperature, the system invoked multi-source knowledge from equipment files, historical maintenance records, and repair manuals to generate a diagnostic report, plan repair steps, and directly dispatch tasks to a third-party work order system.
Discussing the system's strategic positioning, Mr. Zhao believes future enterprise organizational structures will change. Some work will be managed by human resources, while other parts will be overseen by IT departments, whose responsibilities will shift towards data ingestion, model training, and skill generation. Besides hiring human staff, companies can supplement their workforce with AI digital employees configured through computing power. He views Deepexi OS as the core infrastructure for next-generation enterprise IT construction, essentially a management platform for enterprise AI resources and AI employees.
Another key agenda item was the launch of the Embodied Intelligent Brain joint laboratory with Tianjin University. Professor Feng Wei, Dean of the School of Computer Science and Technology at Tianjin University, stated in a speech that the collaboration will focus on three areas: data simulation and synthesis, large model lightweighting, and model inference architecture optimization.
Mr. Yang further explained that this initiative marks DEEPEXI TECH's expansion from the digital world into the physical world. "An AI employee combined with an embodied robot constitutes an intelligent employee in the physical world," he said. He illustrated that ideal embodied intelligence is not just a robotic arm performing movements, but understanding the processes, structures, and production logic behind samples, becoming a business expert on the production line.
Comments