At the Internet Finance Center in Beijing’s Haidian district, Math Magic, the company founded by Ren Lifeng, operates from three locations. One of them is not an office, but a laboratory.
Inside, commercially available 3D printers line the worktables. Every day, the machines print objects generated by the company’s proprietary artificial intelligence model. The team uses the lab to see whether Hi3D-generated designs develop broken surfaces or other defects when printed, then works out how to fix them.
By mid-August, Math Magic’s self-developed Hi3D V3.0 model was entering its final stage before public release. As the team stepped up daily testing and discussions, the volume of test prints surged. Resin waste accumulated faster than it could be cleared, with some of it piled into cardboard boxes in a corner of the lab.
Ren founded Math Magic in early 2024, beginning the next stage of his career after leaving ByteDance. Amid growing interest in large AI models in China, he chose a narrower path: combining AI-generated 3D with manufacturing.
Over the next two years, the company built its own factory in the Pearl River Delta. Its 3D generative AI foundation model is intended to help turn ideas and design drawings into manufacturable objects.
Investor interest followed quickly. Soon after Math Magic was founded, HSG and IDG Capital expressed interest in backing the company.
Ren brought a resume closely tied to several of ByteDance’s major products. In the fall of 2016, he was a core member of the team incubating A.me. Three months later, the product was renamed Douyin. He later headed Xigua Video before becoming a vice president at Pico, where he was responsible for its content ecosystem.
In the second half of 2023, Ren gradually stepped down from several roles within Douyin Group. He formally left to start a company in early 2024.
But when HSG and IDG expressed interest in investing, Ren asked them not to rush.
He told the investors to put the terms and business plan aside and first spend two days touring factories with him in the Pearl River Delta. Representatives from both firms went with him to Panyu, Guangzhou, where they stood in front of production lines in the local jewelry manufacturing cluster.
The exercise was intended to shift the discussion away from Ren’s profile and toward the business itself. He wanted prospective investors to see the manufacturing processes firsthand before evaluating the company’s plan.
Ren acknowledges that his profile can attract capital and talent. But he says he would rather have investors and employees commit because they believe in the underlying business and understand its risks.
On August 20, Math Magic formally released its next-generation Hi3D V3.0 model, which the company said had reached a resolution of 2,048-cube voxels. It also announced that it had raised nearly USD 50 million in Series A+ funding.
Over the previous six months, the company had completed two consecutive funding rounds led by BAI Capital and Hua Capital, with participation from V Fund and existing investors HSG, IDG Capital, DragonBall Capital, Crystal Stream, and Jinqiu Fund.
Two and a half years after its founding, Math Magic has built a factory in the Pearl River Delta and reached the point where its model development directly intersects with manufacturing.
Ren’s previous career centered on consumer internet products including Douyin, Xigua Video, and Pico. Now, much of his attention is on factories, materials, production processes, and the data generated along the way.
What is he looking for there, and where does he think this path will lead?
The following transcript has been edited and consolidated for brevity and clarity.
36Kr: What were you thinking about doing when you decided to leave ByteDance?
Ren Lifeng (RL); To be frank, I wasn’t planning to start a company then.
During my final stretch before leaving ByteDance, I was on vacation. Out of personal interest, I went to look at some industrial clusters, and that was when an idea began to form, somewhat hazily.
After doing some preliminary validation and deciding there might be an opportunity, I chose to go out and build it.
At the time, I knew what I was about to do would have to involve two areas.
One was the supply chain. I needed to get very deeply involved in the supply chain. The other was 3D.
Within ByteDance’s business system, images and video are core media formats. But what I wanted to do was 3D, and especially to combine 3D with manufacturing. So I felt this direction should be pursued independently.
36Kr: What is Math Magic trying to build?
RL: What we want to do is combine 3D generative AI models with manufacturing.
What I did at Douyin was provide a tool for sculpting in the dimension of time. Today, with 3D generative AI, I am providing a tool for sculpting in the dimension of space.
People have eyes and mouths because they need to express themselves. They also have hands because they want to participate in reshaping the world.
What I am doing now is taking the vision we once had of helping users sculpt time and extending it toward helping users sculpt the three-dimensional world.
36Kr: What first led you to think about 3D foundation models?
RL: The idea actually came more than a year before the company.
At the end of 2022, when I had just started overseeing Pico, AI-generated images took off.
My immediate reaction was: if images can be generated, and text can be generated, then audio and video can probably be generated too. So when will AI-generated 3D arrive?
I very seriously went to the algorithm scientists in the AI lab and talked extensively with them about that question.
That was because 3D production was an enormous pain point for virtual and augmented reality.
At the time, if we wanted to produce a piece of interactive 3D audiovisual content, the production cycle was at least two months. You would have 50 people working for two months just on the art pipeline, and that was already considered very fast.
For a traditional AAA game, a two-year development cycle is the bare minimum, and much of those two years is spent on the art pipeline.
If the content ecosystem is not complete or rich enough, you cannot sustain an immersive experience delivered through something people wear on their bodies.
If 3D assets cannot be produced efficiently and with a low barrier to entry, that use case cannot take off.
36Kr: Why did you move from 3D models into physical manufacturing?
RL: I visited a very large number of industrial clusters.
After seeing enough of them, I started wondering whether it might be possible to learn from factory data and build a 3D foundation model that could be integrated with manufacturing.
Building a model comes down to two things: data and a parameterized model.
The most important thing is having a large-scale training corpus, or dataset. To some extent, data determines the ceiling of the model’s performance.
Factories generate huge amounts of data throughout manufacturing. I started asking whether there was data in those processes that we could obtain.
I looked into it, and it turned out there was.
36Kr: Which industrial clusters did you study first?
RL: The first one I took investors to see was the jewelry cluster in Panyu.
From the first day we arrived, we toured factories. I took them through every stage of the production line: what each job was, what tools workers used, what value they created, and which stages of the production process might present a technological opportunity.
Jewelry has a very complete manufacturing chain and deals with a wide variety of materials.
Among rigid materials, you have metals, resins, and gemstones. The forms are extremely diverse, and the processing is difficult enough to be meaningful.
It also deals mainly with hard materials, and equipment for hard materials tends to have a relatively high degree of digitization. More of the processing steps that need to work together are digital.
So by studying this industry, we were getting a window into the broader manufacturing chain. We were able to examine many related supply-chain issues all at once.
36Kr: Why build your own 3D model instead of partnering with another team?
RL: We had actually looked at the models everyone was building at the time.
First, the research was all still very early.
Think about 2023 and 2024. As an independent modality, 3D developed later than images, text, audio, and video. At that point, there were very few models producing genuinely good results.
Second, we did not see any model that integrated particularly well with manufacturing.
The problem we are studying is fairly vertical. A 3D foundation model can be general purpose, but within that generality, its application value is greater in manufacturing.
We also have gaming and film customers using our models, but they use them at specific stages. Manufacturing is different. We go much deeper into the application and provide a series of pipeline-based models and services.
So we chose to build our own model because the problem we needed to solve was fundamentally a production problem.
36Kr: Who were the first investors to approach you, and how quickly did they move?
RL: HSG and IDG were probably among the earliest to express support.
When I finally decided that I was going to start a company, HSG and IDG clearly said they wanted to invest. But I held them back.
I said, don’t rush.
Then I took people from HSG and IDG with me to spend two days touring factories in the Pearl River Delta.
After we finished touring factories on the second day, we held a meeting that afternoon. Only then did I walk them through the business plan.
I told them: look, this business plan is based on the fact that we have now gone through the industrial clusters and seen all of this information. Do you think it is worth investing in?
36Kr: You did not want investors backing you primarily because of your profile?
RL: I don’t deny that people pay more attention to me than they otherwise might, but I don’t enjoy the so-called halo.
When I interview candidates, I especially dread the final stage, when it becomes a two-way conversation.
Someone will say: “Let me tell you the truth. I’m actually coming because of you. I believe you can lead everyone to solve these problems.”
I can understand that, but I’m particularly afraid of hearing it.
I’ve told people many stories that sound wonderful and may contain a compelling vision. But behind them, what I am trying to do carries enormous uncertainty. The cycle is long, the investment is heavy, and the risk is substantial.
A halo cannot cover that up.
So if you ask me how I think about being a high-profile founder, yes, I know it can attract resources and talent. But ideally, what finally tips someone over the line should be that they believe more deeply in the thing itself.
36Kr: You just announced another funding round. Where will the money go?
RL: The largest share will definitely continue to go into foundation model research. That is our base.
First and foremost, we are a technology company. The biggest portion of our R&D spending will continue to go into AI models and related technologies.
The second area is investment in collaboration with industrial clusters.
That includes physical investment in materials, equipment, and processes, as well as connecting our pipeline services with those of partners. We will invest quite clearly in this area as well.
The third area is international expansion.
When you are building something, you need to assemble as many of the factors required for success as possible. Only then do you have a chance of producing a good result.
China has two major advantages. One is its engineering talent dividend. The second is its manufacturing advantage.
If you combine the best engineering talent with the best manufacturing capabilities, then what you build should be global.
36Kr: Hi3D V3.0 reached a stated resolution of 2,048-cube. Why does precision matter so much?
RL: Because in industrial applications, precision represents quality.
Manufacturing has a concept called machining precision.
Different types of equipment are categorized by precision. A standard 3D printer is around 50 microns. Laser equipment is around 20–30 microns. A precision CNC (computer numerical control) lathe is around 10 microns.
If you want to adapt to these types of equipment in manufacturing and demonstrate your product’s value, precision is unavoidable.
Digital assets do not need to become physical objects. But production means putting something tangible into a user’s hands and solving a real problem. If the precision is not high enough, users cannot feel the value.
With this version of the model, we made a major leap from 1,536-cube to 2,048-cube. We can say responsibly that, based on assessments by both our internal team and senior modeling experts we hired externally, it has reached the limit of human modeling precision.
In fact, the machining precision of a lot of equipment is still below the precision of the model. But we are not going to increase precision without limit. We also need to consider applicability.
Going forward, precision may not necessarily be the next thing we have to pursue aggressively.
36Kr: Why not keep pushing precision higher?
RL: At the next stage, improving it even a little more would require an enormous amount of resources, while the corresponding gain would be relatively small.
36Kr: If precision is no longer the main focus, what comes next?
RL: We have three relatively clear directions.
The first is to evolve from integrated forms toward structured ones, breaking a complete object into components and moving into internal structures.
The second is to move from mesh-based representation to parametric representation. Meshes offer relatively little freedom to edit, while parametric models offer much greater controllability.
The third is to move from working with rigid and hard materials toward researching flexible materials and flexible-material processes.
Those are the three paths we are relatively certain about right now.
36Kr: How many more companies can the 3D AI model market accommodate?
RL: I think the industry is still in an early stage of rapid development, and there is plenty of room for incremental growth.
3D is an independent modality, and compared with images, text, audio, and video, it developed the latest.
There is one thing I think we should genuinely be proud of: across all these AI model modalities, 3D is probably the only field in which Chinese teams have remained at the forefront from the emergence of the technology up to today.
Looking at the market, we can see fairly clearly that annual shipments of 3D printing hardware will exceed 10 million units.
That is comparable with annual VR shipments at the time, which were roughly 12 million and 14 million units in 2022 and 2023, respectively.
Annual compound growth in 3D printing-related hardware is around 23–26%, so shipments exceeding ten million units should be highly likely.
Beyond that, many of the directions we are pursuing will move into industrial applications, providing solutions for vertical industries. Those are markets worth trillions of USD, so the opportunity is even larger.
36Kr: What do you think about potential competition?
RL: It’s fine. Looking at use cases, 3D will be used in gaming, manufacturing, film and television, industrial design, architecture, e-commerce, and education. The two biggest areas right now are gaming and manufacturing.
We are also going deep into manufacturing. During training, our model has learned from large amounts of manufacturing data. That is one of our advantages.
Every field within 3D is itself a vertical. The market is large, and doing any one of those fields well requires an enormous amount of effort.
Once you go deep into the industrial chain, the differences between the paths people choose become much more obvious.
Games need low-polygon models and clean topology. Manufacturing needs extremely high precision and integration with processes at the production level.
Even within the same broad field, the objectives pursued by those two applications often point in different directions.
Trying to extend a single foundation model and use it to deeply solve the problems of two verticals this large is probably unrealistic, unless you build separate large teams for each one.
So it is difficult to call that competition.
36Kr: Why does an AI model company need to operate its own factory?
RL: We built our own factory because, in reality, we were forced to.
The fact is that it is very difficult to persuade a normally operating factory to cooperate with you on experiments.
If you stop the production line, orders get delayed. Workers still have to be paid. And most of the results from the trial and error do not belong to the factory anyway.
If I put myself in their shoes and imagine that I were the plant manager, I wouldn’t want to do it either.
Because we could not buy that kind of cooperation, we had to do it ourselves.
I’m not joking. I really am serving as a plant manager.
36Kr: What did operating a factory add to the model development process?
RL: We were in a meeting discussing how the Hi3D V3.0 model was performing in actual printing and which areas still seemed problematic.
Several groups of people were sitting around the conference table. One group had come from Guangzhou to conduct experiments. Another consisted of external hands-on experts. Then there were people from the algorithm and product teams.
Why does a print have broken surfaces? Why are there holes? Why does the system report an error when reading a file?
When we discuss these issues in that setting, we can often get answers from different levels very efficiently.
Once, when we were troubleshooting a problem, someone from the factory very quickly pointed out that it was not a model problem. It was a printer problem.
We actually reported the problem to the printer manufacturer.
After the investigation was completed, we found that the problem really was in the printer’s slicing software.
People who understand hands-on production will also raise issues involving color management.
They will suggest upgrading materials from rigid resin to high-toughness resin, conducting drop tests, and reducing the risk of breakage.
Those are the kinds of questions only people with hands-on experience will raise, and they can help improve the algorithms.
Conversely, once the algorithm team understands a real-world production problem, it can say, “I can solve that for you.”
We need people on both sides to collide intellectually like this.
Traditional manufacturing workers and AI algorithm engineers discussing problems around the same conference table is a scene that I think is probably difficult to find either in China or anywhere in the world.
36Kr: Where did your interest in manufacturing come from?
RL: When I was working in user operations at Baidu, I was responsible for managing Baidu Tieba.
To understand users, you first had to become a user yourself.
Back then, the important thing on Tieba was to really immerse yourself in the communities. So I followed my own interests. I looked at jadeite. I looked at tea.
As early as 2013 and 2014, I was buying jadeite over WeChat from people in the industry whom I had met through Tieba. Even the cheapest pieces cost at least RMB 5,000 (USD 740) each.
At the time, my monthly salary was less than RMB 5,000, so I would save up to buy them.
Why did I dare to buy through WeChat? Because in that Tieba ecosystem, the moment you dared to sell a fake, someone would expose you. So sellers would not behave recklessly.
If you study jadeite, you naturally branch out into other types of jade and jewelry.
If you study tea, you start studying teaware and ceramics.
Following that thread, my interests gradually spread into many consumer products with aesthetic and cultural dimensions.
Jadeite gave me one important insight. Most raw materials are actually very ordinary. But after they are carved and shaped, the perceived value can increase dramatically.
For a while, I kept thinking that perhaps this was exactly where the leverage of technology could create value.
For decades, because of my personal interests, I have often visited these industrial clusters purely as a hobby. I have been to many of them and have become familiar with the people who work behind the scenes.
36Kr: So you were already familiar with these industrial clusters before starting Math Magic?
RL: Right. It wasn’t something I had only started doing in those few days. I had been visiting them for more than a decade.
36Kr: What did you learn from visiting factories?
RL: China does not lack good designers, but it lacks good pattern makers.
That problem exists across consumer goods.
In apparel, a pattern maker is what we call a tailor.
Tailors are highly respected within Europe’s social and professional system.
A tailor receives a designer’s clothing sketch, sometimes nothing more than a hand-drawn pencil concept, and uses memory to translate it into the corresponding fabric materials, pattern pieces, and accessories that need to be sewn onto the garment.
In reality, at many garment factories, a piece of clothing cannot enter mass production until a pattern maker has assembled and finalized the pattern, after which its parameters are collected.
Plush toys are another category that has grown very strongly in recent years. At a medium or large factory producing plush toys, there are very few pattern makers. They typically need at least ten years of experience. The good ones have 15 or 20 years of experience.
At that point, we would call them master craftsmen. They, too, receive a flat concept drawing and translate it into the corresponding pattern pieces.
36Kr: Is that intermediate step where you see technology creating leverage?
RL: That is the simplest way to understand it.
36Kr: You often use the word “democratization.” What does it mean to you?
RL: After working on Douyin, we saw a lot of content we never would have imagined before.
Behind all of it were real, living people. They all had incredibly rich inner worlds and lives.
You realize how extraordinary the world is.
I especially like one of YouTube’s early slogans: “Give everyone a voice and show them the world.”
Give everyone a microphone and let them express the world inside themselves.
In one sense, that was what Douyin was doing: democratizing expression in the digital world.
Overseas, we have seen consumers willing to pay a premium for designers’ creative work. Even if the quality of the product itself appears ordinary, if it reflects your own design ideas and your own craftsmanship, people are willing to pay for it. That is respect for creativity.
So we should respect original work and copyright. We should also provide good services, good tools, and good platforms for people who want to express their creativity.
Ultimately, what we still want is consumer-facing democratization. We do not simply want to give factories and manufacturers efficiency tools.
We want to connect the entire pipeline and make those tools available to ordinary people through low-barrier, platform-based services.
KrASIA features translated and adapted content that was originally published by 36Kr. This article was written by Wen Lihong for 36Kr.
Note: RMB figures are converted to USD at rates of RMB 6.73 = USD 1 based on estimates as of September 10, 2026, unless otherwise stated. USD conversions are presented for ease of reference and may not fully match prevailing exchange rates.