Zhiman raises tens of millions of yuan to make semiconductor companies dozens of times more efficient with LLMs

In 2025, the “AI factory” that NVIDIA founder and CEO Jensen Huang described is taking root in China’s semiconductor industry. A traditional factory turns raw materials into products; an AI factory turns data and compute into intelligent services that help companies cut costs and work faster.

We recently met a start-up called Zhiman that is turning the AI factory into something real by combining vertical models, AI search and agents. Its AI solutions are already in use at several companies, where they have made semiconductor work dozens of times more efficient.

Zhiman was founded in June 2024 and recently announced an angel round of tens of millions of yuan from Oriza Hua (元禾璞华).

Over the past year it has released nine vertical models for semiconductor R&D and production, including a vision model, an on-device model, a code model and a chat model, and built AI search and agents on top of them.

Zhiman’s products

After a year of working with customers, Zhiman’s founder Ben told us, he sees a few basic things semiconductor companies need from AI products:

1. Results must be accurate and traceable to their sources.
2. For data security, nothing that goes in or comes out may leave the company. The product must deploy on-premises and support the company’s own fine-tuning.
3. The product must make a real difference to efficiency in day-to-day work.

To meet these needs, Zhiman started with data. It prepares corpora of semiconductor knowledge, brings each company’s own data into training, and trains for each scenario with SFT and RL, so that every company ends up with a model of its own.

On security, Zhiman works mainly in private deployments: it deploys and trains on the customer’s side, and also offers private all-in-one appliances and clusters. Inside a company, its AI products give each employee viewing, search and Q&A permissions according to their business unit and department.

On efficiency, the question that matters most, Ben told us that Zhiman’s AI agents and AI search make semiconductor teams 10 to 90 times more efficient, depending on the scenario.

In R&D, for example, engineers use Zhiman’s AI search to look up process standards, research papers and patents. The AI summarises the key points, helps them sharpen the question, and lets them check each answer against the original text, which saves much of the time spent collecting and analysing material.

In production, Zhiman has built a series of semiconductor expert agents that work alongside manufacturing software such as MES.

Its wafer defect analysis agent, for instance, connects directly to the customer’s equipment and systems.

When a wafer defect image is captured, the agent uses Zhiman’s vertical models to identify the defect, explain its likely cause and propose a fix, down to the machine parameters to adjust. Engineers can ask the agent follow-up questions. Zhiman offers similar agents for component design and other tasks.

Ben told us that Zhiman already has a working commercial model, with deployments at several enterprise customers, and will keep adding customers and scenarios this year. The company has also opened a new funding round and is hiring.

Chen Yu, managing director at Oriza Hua, which led the round, told us he had been following Zhiman since the second half of 2024.

Oriza Hua is a well-known investor that has spent ten years in semiconductors and backed more than 200 start-ups. With the arrival of the AI era, the firm says, it has moved quickly into AI, including large models, embodied intelligence and AI applications.

Chen Yu told us that Zhiman is an important part of Oriza Hua’s bet on vertical B2B applications of large models. From managing production knowledge, to tracing and answering problems, to agents working in production, Zhiman’s models understand the semiconductor industry better than general-purpose models such as DeepSeek.

During and after the investment, Oriza Hua has introduced Zhiman to several leading Chinese semiconductor manufacturers, and deployments have begun in a number of scenarios. Oriza Hua believes Zhiman will play a key role in rebuilding China’s AI factories around intelligence.

What follows is an excerpt from a conversation between 四木相对论 and Zhiman’s founder, Ben.

Q1: First, please introduce Zhiman and its products.

Zhiman: We are a technology company focused on vertical large models for the broader semiconductor industry. We offer a complete enterprise AI solution, including foundation models, AI search and semiconductor expert agents.

We already have a fairly broad range of enterprise AI products. In R&D, for example, we provide AI search. Engineers can quickly look up R&D material such as process standards, research papers and patents, have the AI summarise the key points and sharpen the question, and trace every answer back to the original text. That saves a great deal of the time spent collecting and analysing material.

In production, we provide more specialised agents, for example expert agents for yield improvement and defect analysis. These agents work with manufacturing software such as MES to support production. By combining on-device models for semiconductor equipment with software MCP, our agents give semiconductor production equipment something like self-driving software.

Q2: Why did you choose the semiconductor industry?

Zhiman: Our core team was using AI to help companies cut costs and work faster before ChatGPT appeared. That experience taught us that AI can do this in many scenarios, and large models can help the semiconductor industry become more efficient and break through the bottlenecks where China depends on others.

B2B companies also have clear needs and clear reasons to pay, and the industry is already fairly digitised, which makes it a good place for a start-up to begin. In the end, we chose semiconductors.

Semiconductor scenarios also have a lot in common with parts of new energy and aerospace, and we will gradually expand into those adjacent industries.

Q3: What do semiconductor companies uniquely require from large-model products?

Zhiman: Vertical models for semiconductors have a very high bar, and semiconductor companies ask for things general-purpose models struggle to deliver. Their basic requirements are:

1. Results must be accurate and traceable, with no hallucinations.
2. Data security: nothing may leave the company. The product must deploy on-premises and support the company’s own fine-tuning.
3. It must make a real difference to efficiency and solve real problems.

To meet these requirements, we first trained on a large corpus of semiconductor knowledge, and then brought each company’s own data into fine-tuning, so that it ends up with a model of its own.

On security, our standard products and solutions are mainly delivered as private deployments. Inside a company, our AI system gives each employee viewing, search and Q&A permissions according to their business unit and department.

On efficiency, we have different products for R&D and for production, and we give companies a lot of freedom: they can call our APIs and use our fine-tuning services. The speed-up is significant. Based on our evaluations, our products make teams 10 to 90 times more efficient.

Zhiman’s core model architecture

Q4: How do Zhiman’s search and agents differ in use? What problems do they solve?

Zhiman: Search suits R&D knowledge work: looking up process standards, searching patents, reading research papers.

Agents carry out more specialised work. In scenarios such as yield improvement or simulation, an agent uses the industry model we trained, together with the company’s existing tools such as simulation software and EDA, to get the task done.

In other words, agents go deeper into manufacturing. They support semiconductor production, for example by proposing a plan to improve yield. Until now, details like how to raise yield or how to set a parameter depended on people’s experience.

In design, an engineer can enter a requirement and the product produces the drawing automatically. That makes a company at least dozens of times more efficient.

Compared with the AI the semiconductor industry used before: in inspection, earlier AI relied on high-precision image recognition. Once it found an error it raised an alarm, and that was mostly the end of its job. Zhiman’s AI identifies and analyses the problem automatically, and also tells the engineer how to fix it.

Bringing vertical models, AI search and expert agents into production is transformative.

In some scenarios, our AI system finishes in ten minutes work that might have taken a company six months. Where teams once needed thousands of physical experiments to get one real result, with AI, or with AI tuning the parameters, they may need only a handful or a few dozen, and sometimes none at all. That changes how these companies work.

An industry that can scale

Q5: How do today’s large-model products differ from traditional software such as SaaS?

Zhiman: First, they occupy a different niche. With traditional software, including SaaS, people adapt to the software’s rules. AI adapts itself to the production logic and its rules. It understands both ways: how the software works and how the person using it works. It can also draw on historical data and operate other software to complete a task.

Second, large-model products carry much less customisation. A model is, at heart, a neural network trained on multimodal data, so once the data is in place, crossing into a new industry is relatively easy. Large models lower the barrier to taking a product from one industry to another.

Third, vendors used to write code by hand to meet each customer’s custom needs within an industry. Large models adapt on their own: moving to a new customer only means personalising with that customer’s data, while the industry knowledge stays. A large model can grow an enterprise model out of an industry model, and keep improving itself from feedback. That lowers the barrier to serving different customers in the same industry.

One of the biggest results is that large-model products can grow faster than enterprise software.

Q6: How strong is demand among semiconductor companies today, and how willing are they to pay?

Zhiman: We have talked with many senior leaders in the semiconductor industry, and they saw the value of large models early. They are very willing to try new things and bring a great deal of creativity and energy. They don’t just follow the leaders; they look for new ways to strengthen their products and their companies.

Often it is leadership that drives adoption further down, so the whole company pulls together to learn and use AI.

These companies are very willing to pay for AI. They have already seen the efficiency gains, and as long as data and products are secure, they are happy to try new AI products, and the share of budget they put into AI keeps rising. Everyone wants to keep up with the frontier.

Q7: The semiconductor industry has many segments and many kinds of companies. Does that make it harder to win customers?

Zhiman: I think different semiconductor companies still share common scenarios. Every design company, for example, wants to optimise its routing. Whether the optimisation is on the die or on the circuit, a multimodal large model can handle it; only the training data differs.

Q8: Will Zhiman move into industries beyond semiconductors?

Zhiman: Large models are changing how the semiconductor industry produces, and there is still a great deal for us to do here. Going forward, we will decide what to take on next by looking at each industry’s production logic, whether its data is complete, and whether we can truly help it upgrade.

Expanding into adjacent industries such as aerospace and new energy manufacturing is part of our plan. The logic is that semiconductors sit at the top of the advanced-manufacturing pyramid and make the toughest demands on their suppliers. If we can crack this field first, we will have a big advantage moving into others.

We also see that high-end manufacturing, semiconductors included, badly needs people who truly understand both the industry and AI.

To support the growth of the business, we have opened a new funding round and keep hiring outstanding people.