HSG + Shunwei + ZhenFund and other investors are betting on world models

New financing dynamics have emerged in the field of World Models.

Recently, LiberAI, a world model development company, announced the completion of hundreds of millions of yuan in Series Pre - A financing. The lead investor is Shunwei Capital, and follow - on investors include Cathay Capital, Yuanhe Origin, Mufar Science and Technology Innovation, etc. Existing shareholders Sequoia China and ZhenFund have continued to increase their investments.

LiberAI stated that the funds from this round will be used for the technological iteration of physical world models, the implementation of embodied intelligence products, the recruitment of core talents, and the construction of computing power systems to promote the commercialization of technologies.

It is worth mentioning that this round of financing is the third financing completed by the company this year, only one month after its previous round of financing. In May this year, LiberAI announced the completion of approximately 500 million yuan in Angel + round financing, jointly invested by ZhenFund, Sequoia China, Meituan Longzhu, Shunwei Capital, etc. Among them, since ZhenFund's exclusive investment in the seed round in 2025, it has increased its investment in three rounds of financing in total.

The founder is from Tsinghua University and is a member of the post - 2000 generation

The world model was originally a concept in the field of reinforcement learning. It refers to an AI agent constructing an internal representation of the external world by learning the generative model of the environment, so that it can simulate and test action plans in its "mind".

Statistics show that currently, the mainstream technical routes of world models in the academic field include: observational generative models, which excel in "realism"; latent space models, which are strong in "efficiency"; reinforcement learning - oriented models, which are strong in "decision - making"; and object - centered models, which are strong in "interpretability". They explore the ways for AI to understand the physical world from different dimensions and also correspond to different technical challenges and application scenarios.

Looking back at LiberAI, the company was established in December 2025. Its affiliated company is Beijing Jiangxian Technology Co., Ltd., with a registered capital of 1.4604 million yuan. The legal representative is Liu Songming.

According to its official information, LiberAI focuses on the research and development of physical world models and embodied intelligence. Through the technical layout of "human UMI data + world model", it builds a full - stack technical system for physical intelligence. Its core products include large embodied intelligence models, supporting UMI hardware, and data collection and training systems, which can be applied in fields such as robotics, autonomous driving, and industrial simulation.

LiberAI's founder, Liu Songming, is a member of the post - 2000 generation. When he embarked on entrepreneurship at the end of 2025, he had just graduated from the Department of Computer Science at Tsinghua University and was only 23 years old. During his undergraduate studies, he won the Special Scholarship of Tsinghua University, ranked first in his major, studied under Professor Zhu Jun, a scholar in the field of video generation models, and published many papers as the first author in top - tier AI conferences such as ICML and NeurIPS.

In 2024, Liu Songming switched from the field of AI for Physics to embodied intelligence, led the RDT series of research, and released the world's first base model RDT - 1B using the large - scale pre - training + diffusion Transformer paradigm; the following year, he released the first embodied model RDT - 2 using the large - scale UMI un - embodied human data pre - training paradigm.

At the end of 2025, Liu Songming founded LiberAI. Four months later, he led the team to launch the world's first high - degree - of - freedom dexterous operation embodied basic model, which can be equipped with a five - finger dexterous hand to complete delicate operations such as peeling bananas with both hands, unscrewing bottle caps, and stir - frying in a pan. The company's team spans top - tier laboratories at Tsinghua University and hardware teams of leading companies, with the ability to build models, conduct training, and coordinate software and hardware.

In terms of technical route, LiberAI chooses pre - training of the original physical modality, rather than mainstream directions such as physical engines or video generation. It is reported that its core innovation lies in modality alignment and inductive bias. Modality alignment refers to using the generalization of massive video data to support the physical modality with scarce data; inductive bias is to inject physical laws as human prior knowledge into the model.

LiberAI believes that the core reason for the insufficient generalization of current embodied models is the lack of causal modeling, that is, the model cannot understand that "force is the cause and the state of motion is the effect", and cannot complete long - range tasks like humans by perceiving force.

In terms of business progress, LiberAI said that it plans to conduct scenario - based joint verification with leading robotics companies and autonomous driving companies in the next year to explore the actual effectiveness of physical world models in complex manipulation, path planning, and environmental prediction.

ZhenFund, Sequoia China, etc. continue to increase their investments

In terms of equity structure, industrial and commercial information shows that Liu Songming, the founder of LiberAI, currently holds a 28.3736% stake, making him the largest shareholder and the actual controller.

Looking at its financing history, reporters from the Science and Technology Innovation Board Daily noticed that since the company was established in December 2025, LiberAI has completed four rounds of financing in a dense manner within half a year.

Specifically, in December 2025, ZhenFund made an exclusive investment in its seed round; in March 2026, Sequoia China increased its investment in its angel round; only two months later, LiberAI completed approximately 500 million yuan in Angel + round financing, jointly invested by ZhenFund, Sequoia China, Shunwei Capital, and Meituan Longzhu.

The company raised hundreds of millions of yuan in this round of Series Pre - A financing, led by Shunwei Capital, with follow - on investments from Yuanhe Origin, Cathay Capital, Mufar Science and Technology Innovation. Existing shareholders ZhenFund and Sequoia China have both continued to increase their investments.

Overall, since its exclusive investment in the seed round, ZhenFund has increased its investment in three rounds of financing; Sequoia China entered in the angel round and has invested in three consecutive rounds; after making its first investment in the Angel + round, Shunwei Capital quickly became the lead investor in the Series Pre - A round.

The world model field is gradually heating up

World models are accelerating from an academic concept to industry development.

Research by CITIC Securities shows that AI giants such as NVIDIA, Google, and OpenAI have all regarded world models as a potential route to the next generation of intelligence. Coupled with the recent intensive financing of academic giants such as Fei - Fei Li and Yann LeCun, the world model track is expected to achieve double dividends through model iteration and valuation increase.

As of now, global world models are mainly divided into three mainstream directions: prediction - driven, content - generation, and action - driven.

Among them, OpenAI's video - generation school, represented by Sora, relies on large - scale video data to enable the model to master physical laws; World Labs founded by Fei - Fei Li follows the 3D spatial intelligence route; Yann LeCun's AMI Labs bets on abstract representation models and causal inference.

Meanwhile, capital is pouring into the world model field at an accelerating pace. In February this year, Fei - Fei Li's World Labs completed a new round of $1 billion in financing; in March this year, AMI, a world model company founded by Yann LeCun, completed a $1.03 billion seed - round financing, setting a new record for the seed - round scale in the European AI field.

Looking at the current domestic competitive landscape, the world model field is still in the early stage of development, with diverse and differentiated technical routes. Capital is flowing into leading startups at an accelerating pace, and commercialization is mainly focused on scenario verification and early orders, with no absolute leader yet.

In addition to LiberAI, Xingyuanzhi, incubated by the Beijing Academy of Artificial Intelligence, follows a full - stack route of "hardware - software integration and edge - side deployment", has raised a total of 1 billion yuan, and has made early progress in commercialization with orders; Inverse Matrix Technology adopts a technical path of reinforcement learning and latent - space causal modeling, aiming to build a "physically correct" general world base model and has completed angel - round financing of over ten million dollars.

From the policy perspective, world models have gradually come into the view of industrial planning. In December 2025, the World Model Technology Report led by the China Academy of Information and Communications Technology was successfully approved by the International Telecommunication Union, marking that relevant research on world models has been officially included in the international standardization work in the field of artificial intelligence.

In terms of market size, according to Morgan Stanley's prediction, by 2035, the industrial scale empowered by world models will reach $10 trillion; according to Gartner's prediction, the global market size related to world models will exceed $5 billion in 2026.

This article is from the WeChat official account "Venture Capital Daily". Authors: Dai Jiayi and Li Yu. Republished by 36Kr with permission.