36Kr Exclusive | Incubated by Dongguan Institute of Artificial Intelligence, targeting the R&D of physical AI in the mold manufacturing field, this enterprise has obtained tens of millions of RMB in seed round financing from Sinovation Ventures.

Written by Qiao Yujie

Edited by Yuan Silai

This article is 2100 words long, with an estimated reading time of 5 minutes.

Hard Krypton learned that Hemu Artificial Intelligence Technology (Chengdu) Co., Ltd. (hereinafter referred to as "Hemu AI"), a physical AI manufacturing company, has recently completed a multi-million-yuan seed round of financing, with Sinovation Ventures as the investor.

Hemu AI was founded in May 2026, incubated by Dongguan New Generation Artificial Intelligence Industrial Technology Research Institute in cooperation with mold manufacturing enterprise Duanpin Precision. It is a company focused on physical AI for molds, manufacturing semantic computing and industrial implementation.

Specifically, Hemu AI hopes to further learn and replicate the process experience accumulated by veteran technicians on the basis of deep learning of mechanical drawings, build an AI process brain for mold manufacturing enterprises, and extend its capabilities to product quotation, mold design, DFM, process scheduling, production processing, quality control and inspection and other links, so as to improve the automation and intelligence level of the entire manufacturing process. 

(Source: the company)

The company's founder Xu Chen is the founding dean of Dongguan New Generation Artificial Intelligence Industrial Technology Research Institute, with more than 15 years of in-depth experience in the fields of intelligent manufacturing and industrial AI. Co-founder Zou Luhua has been engaged in the high-end precision mold industry for nearly 20 years and is an expert in high-end mold manufacturing management. 

Choosing mold manufacturing as the entry scenario is related to the long-standing "veteran technician dependency" problem in the industry. At present, labor costs in mold manufacturing account for about 70% of the total variable costs. On the one hand, there is a shortage of senior skilled workers in the manufacturing industry, and rising salaries have further compressed the profit margins of enterprises; on the other hand, key positions in mold manufacturing generally require practitioners to have more than 3 to 5 years of practical experience, resulting in a long talent training cycle and high threshold, and the industry is currently facing a gap between the outgoing skilled workforce and insufficient new qualified talents.

The Hemu AI team has conducted comprehensive research and investigation on industrial manufacturing enterprises since 2024. Xu Chen told Hard Krypton: "We have visited more than 200 enterprises and found that factories basically do not allow data to be connected to the Internet due to core confidentiality management needs. Many enterprises cannot directly deploy large models at all. Therefore, if we want to develop industrial AI, we must first solve the problem of localized computing, and the collaborative mode of large and small models is the only feasible implementation path."

To solve this problem, the team spent two years sorting out 100,000-level high-quality drawing data, and built an industry knowledge system by mining the experience of more than 100 veteran mold technicians. According to the company, the team has initially completed the construction of the semantic space and computing base for precision mold manufacturing based on the collaboration of large and small models, and has completed the development of nearly 10 agents. Some of the agents already have experience equivalent to that of a mold engineer with more than 5 years of working experience.

In terms of specific applications, Hemu AI has currently applied its AI capabilities to process programming, intelligent quotation and process route planning and other links.

(Source: the company)

For example, in the process programming link, AI can identify part features, automatically select different processing methods such as CNC, electric discharge and grinding machine, and optimize processing costs while meeting accuracy requirements. At present, the processing efficiency of this link has reached about 6 times that of manual work. In the quotation link, the system can calculate the cost in reverse based on its understanding of the manufacturing process to achieve accurate quotation, while reducing the labor input of traditional quotation positions. In terms of process route planning, AI can participate in process scheduling, recommend corresponding processing paths according to processing conditions, and improve the overall production scheduling efficiency.

In terms of technical path, Hemu AI adopts the architecture of "large model base + small model cluster + process knowledge precipitation". Among them, the large model is mainly responsible for understanding scenario semantics and disassembling tasks, while the small models are responsible for executing specific processes. Aiming at the features that appear frequently in the manufacturing process, such as geometric features of holes and cavities, the company trains dedicated small models to achieve higher-precision process execution.

At present, the team has trained more than 1,600 sets of small models for high-precision task execution in different manufacturing scenarios; the process knowledge system further constrains and verifies the output of the models to make the results conform to the actual manufacturing conditions. Based on this technical architecture, Hemu AI has currently developed agents such as 2D drawing error correction, PDF to CAD, 2D to 3D, and can complete process preprocessing and autonomous programming in 15 types of processes including CNC, electric discharge and wire cutting.

In terms of commercialization, Hemu AI currently mainly provides FDE accompanying AI empowerment services, as well as standardized software and hardware products of agents. Its goal is to help mold manufacturing enterprises shorten the proofing process, reduce the cost of manufacturability evaluation, reduce mold change time, and improve the first-pass yield in the large-scale manufacturing process.

In the future, Hemu AI also plans to further develop hardware products around the existing AI brain, embed semantic computing capabilities into localized equipment, intelligent inspection equipment, robots and other industrial carriers, and gradually form a "brain + carrier" product sales model.

The following is an excerpt from the communication between Hard Krypton and Xu Chen, founder of Hemu AI:

Hard Krypton: What is the difference between Hemu AI's Uniwood large model and the current industrial software?

Xu Chen: Industrial software is a process operation tool that assists humans, and a bridge connecting humans with processing objects and manufacturing equipment. However, industrial software cannot understand the manufacturing principle, and still requires veteran technicians to learn to operate and use tools to complete industrial tasks. The self-developed Uniwood large model of Hemu AI is a semantic computing system for precision mold manufacturing driven by physical constraints, mechanics and assembly principles, topology and structural analysis, material properties and mechanics. It can complete manufacturing tasks independently and has autonomous evolution capabilities.

Hard Krypton: What is the difference between the construction path of Uniwood and the current industrial Agent?

Xu Chen: The current development direction of industrial Agent is more focused on improving the tool operation efficiency of veteran technicians. For example, there are many current applications that use language commands to drive parametric design of parts, and load part of fixed experience and constraints in the process. This is an advanced stage of "assisted driving", which can greatly improve manufacturing efficiency. But Hemu AI's Uniwood is committed to the understanding of physical AI for manufacturing, and autonomous evolution is its core label. Only by using its comprehension to control more and more powerful industrial software can it become a real "highly intelligent driving".

Hard Krypton: How is the experience of veteran technicians extracted? How do the large and small models collaborate with each other?

Xu Chen: After long-term on-site research in factories, we have initially completed the closed loop of token computing in the field of precision mold manufacturing. Veteran technicians use chain of thought to drive a target task. As long as their chain of thought can be accurately decomposed into standard SOP, the experience can be represented by algorithms in segments. In actual operation, after the Uniwood large model understands the scenario semantics, it will dispatch thousands of expert small models to repeatedly adapt to the scenario, so as to approach the deterministic manufacturing target. At present, after 5-8 rounds of self-evolution, the self-developed "Little Crayfish" based on Uniwood can achieve an accuracy rate of over 95% in the recognition of some key processes, which is fully comparable to the capabilities of veteran technicians.