A post-2000s alumnus from Zhejiang University who shares the same mentor as Liang Wenfeng is developing world models and plans to raise another 1 billion yuan in financing | 36Kr Exclusive

Written by Wang Xinyi

Edited by Zhang Yuxin

The world model track is about to embrace another large-scale financing deal.

Exclusive information from Intelligent Emergence shows that Moxin Technology, a firm focusing on the 4D world model sector, is about to close a new round of financing, with the financing scale expected to reach 1 billion yuan, and its valuation is likely to jump to nearly 10 billion yuan. This will be its fifth round of financing this year, as well as the second round in the past nearly one month.

It is learned that after the completion of the previous round of financing, the valuation of Moxin Technology has approached 4 billion yuan, which means that after the new round of financing is completed, its valuation will grow to about 2.5 times within only one month.

The upcoming new round of financing for Moxin Technology adds new momentum to the world model track that seems to have gradually cooled down. According to sources, Moxin Technology has completed the shareholding system reform recently, and the new round of financing continues the previous feature of intensive participation of industrial capital. Its historical shareholders include SZVC, Zhejiang Venture Capital, Hubble of Huawei, Legend Holdings, Fuhan Micro, JinkoSolar, SMIC PE, JH Semi, Prosperity7 Ventures, CDH Investments Fuyi, Loongson Venture Capital and other institutions.

Chen Tianrun, founder and CEO of the company, is a post-2000s generation, and also a 2022 PhD student at the College of Computer Science and Technology of Zhejiang University. He studied under Pan Yunhe, academician of the Chinese Academy of Engineering and former president of Zhejiang University, and is a fellow apprentice of Liang Wenfeng, founder of DeepSeek. His research directions include 3D vision, spatial intelligence and world model algorithms, and he has published more than 50 papers in journals such as Science and Nature Photonics, as well as conferences including ICCV and CVPR.

Founded in 2021, Moxin Technology started its business with 3D printing hardware. It was initially positioned to develop consumer-grade printing hardware, mainly targeting families, students and creative enthusiasts, with a cumulative revenue of tens of millions of yuan.

Years of accumulation in the 3D field made Chen Tianrun realize that 3D AI modeling has its own Scaling Law, so a bolder decision took shape gradually. At the end of 2024, he decided to lead the team to shift to spatial intelligence, focusing on 3D scene modeling and world models. "Whether it is robots or autonomous driving, a 'brain' that can understand and simulate the physical world is ultimately needed, and what we are making is exactly this 'brain'," Chen Tianrun mentioned.

The team soon delivered remarkable results. In December 2025, Moxin Technology released the prototype of the interactive world model KOKONI-World. In July this year, Moxin, together with Academician Pan Yunhe's team and Huawei Cloud, released the world's first Flash World Model — MoWorld, which has 28 billion parameters and adopts the MoE (Mixture of Experts) architecture. This is a high-frame-rate interactive world model. Users can input through images, texts and other methods to generate complete high-definition 3D scenes that conform to real physical logic, support 6-degree-of-freedom (6-DoF) camera control, and support generating videos up to 2000 frames long.

△ Based on one photo or a video, MoWorld can construct a 4D model that simulates the real physical world. Source: the enterprise

The high frame rate is reflected in its ability to realize real-time 3D generation at a speed of 50 frames per second; real-time interaction means that users can control the perspective and actions in real time through keyboards, mice and other devices. After receiving the input, the model immediately generates the next frame of image, and maintains the consistency of spatial structure, object position and historical state.

As mentioned in the MoWorld paper, in the fields of computer graphics and interactive media, 30FPS is widely regarded as the minimum threshold for realizing real-time rendering and providing a smooth interactive experience. The frame rate of 50FPS means that each frame only takes 20 milliseconds to generate, which can quickly respond to user instructions, and is especially important in scenarios that are very sensitive to latency such as embodied intelligence and autonomous driving.

The world model of Moxin Technology does not need to rely on high-end GPUs. The full stack is designed based on domestic NPU, and a framework from data generation, pre-training, distillation to efficient inference is constructed. This model runs entirely on the Ascend NPU super node. In terms of inference cost, its solution can reduce the cost by about 70% compared with the GPU solution of the same scale, which is only 30%-50% of the cost of existing world models.

The world model has almost become a must-contested area for major manufacturers and startups at home and abroad this year. Tech giants including NVIDIA, Google, Tencent, Alibaba, ByteDance have launched or are working on launching their own world models; hot capital has also successively poured into world model startups. Jiajing Vision has completed three rounds of financing of more than 1 billion yuan this year, while Liber AI has completed 5 consecutive rounds of financing in 8 months.

Riding on the trend of the world model, Moxin Technology has quickly closed round after round of financing. 20 days before the upcoming new round of financing at the 1-billion-yuan level, it just disclosed the previous round of 500-million-yuan level financing participated by SZVC, Zhejiang Venture Capital, Hubble of Huawei, Legend Holdings and other institutions. Before that, it also completed three rounds of financing at the 100-million-yuan level: the nearly 100-million-yuan Pre-A+ round financing participated by Hubble of Huawei in March, the nearly 100-million-yuan strategic round financing participated by Zhejiang Venture Capital, Legend Star and other institutions in May, and the several-hundred-million-yuan Series A round financing completed in June, with Hubble of Huawei participating in the investment.

At the same time, the paths to realize world models are also different. The school represented by Fei-Fei Li believes that the core of the world model is spatial intelligence, and it is necessary to construct a measurable, editable and accessible explicit 3D structure, rather than just generating realistic images; while the school represented by Yann LeCun chooses the direction of latent space prediction, and proposes JEPA (Joint Embedding Predictive Architecture), which enables the model to understand visual causality and infer what the state will become at the next moment in a certain abstract space.

Moxin Technology has chosen another route — the implicit structure route, which enables AI not to directly generate image pixels, but to first establish an abstract representation of "3D space + time" internally, and all scenes are generated from this representation. Moreover, it puts more emphasis on high-frame-rate real-time interaction and engineering implementation of full-stack domestic NPU deployment.

The features of low cost and high real-time performance also give Moxin more advantages in cost control and industrial implementation. In industries such as humanoid robots and autonomous driving, the digital simulation training field built based on MoWorld-3D can provide a large number of training scenarios for robot training, allowing robots to complete high-frequency action drills in virtual production lines, home environments and other spaces, providing infrastructure for the large-scale development of embodied intelligence.

After the release of the MoWorld model, Moxin has accelerated the exploration and application implementation in multiple scenarios such as industrial digital twins, embodied intelligence training, urban and rural planning, and cultural tourism consumption.

Moxin Technology has provided a good sample for non-"model-native" companies. However, under the siege of giants and star companies such as NVIDIA, Tencent and World Labs (Fei-Fei Li's startup), whether Moxin can gain a firm foothold, the answer has just begun to be written.