23 New World Model startups emerge in China in 7 months; 2026 sees rapid growth

In mid-June, IT Juzi published a report reviewing domestic startups developing world models, attracting significant market attention.

As this report is being shared, we are discovering new world-model companies being founded and funded every day.

As I said at the end of that report: "World models have just begun—let’s see where this takes us."

Now, given the high level of interest, let’s take another look—

Seven months into 2026, a noteworthy trend has emerged in China's AI startup market:

In the world models赛道, the number of newly established companies continues to rise, reaching 23 as of early August.

What does this number mean? For comparison, 20 world model companies were established in all of 2025, while only 10 were established in 2024. In just seven months of 2026, this number has already surpassed last year’s total.

More intriguing is the attitude of capital.

Of these 23 companies, 18 completed their first or first two funding rounds within months of establishment; two were labeled as unicorns at the seed or Pre-A stage, and four joined the IT Juzi Qianlima club (valued at over RMB 1 billion).

Top-tier institutions such as Sequoia Capital China, Hillhouse Capital, Tencent Investment, Ant Group, Matrix Partners, Sinovation Ventures, and Shunwei Capital have made concentrated investments, while industry players including Xiaomi Strategic Investment, AgiRobot, Looxer Robotics, and StarSea Map have also entered as strategic investors.

On August 3, Breakout Bot announced the completion of its nine-digit Pre-A funding round, becoming the latest standout project in this sector;

On July 31, Quantum Power, founded by Li Qiang, former CTO of Cainiao, completed a seed round exceeding RMB 100 million.

On July 24, Yiseng Brain in Shenzhen completed a tens-of-millions seed round.

The pace of financing has not slowed down; rather, it has accelerated.

The common characteristic of these companies is:

Since day one, we’ve embedded the “World Model” into our company’s core positioning, rather than treating it as an ancillary concept to any single business line.

However, they have different understandings of the world model—

Some are reconstructing 4D spacetime, others are building causal reasoning engines, some are creating the central brain for scientific discovery, others are simulating the dynamics of drug molecules, and some are developing world action models for logistics and warehousing.

The concept of a world model is being simultaneously enriched by a new wave of innovators from various directions.

But more important than the technical roadmap are the people.

The founders of these 23 companies have incredibly diverse backgrounds:

A 22-year-old undergraduate from Yuanpei College, Peking University,

A Tsinghua University associate professor who is the first author of a paper featured on the cover of Nature

Key personnel have left Alibaba's Qwen.

A serial entrepreneur who has previously developed AI chips and commercial rockets.

A NYU PhD who became a doctoral supervisor under the age of 30

With a former CTO of Cainiao bringing logistics expertise to the field,

A chief scientist has left a star company valued at $20 billion.

Their backgrounds form the most fascinating portrait of AI entrepreneurship in China in 2026.

This article categorizes the 23 newly established companies in 2026 into five groups based on their technical approaches, and introduces each group’s team background, technical direction, product status, and funding situation.

Data Rule: IT Juzi Database - Company Database - Tag "World Model"

To build the "operating system" of the physical world

This group of companies aims to build a general-purpose world model foundation, not limited to specific vertical scenarios, similar to foundation models in the field of large language models.

Their technical approaches differ in focus, but their goal is the same: enabling AI to achieve general understanding and simulation capabilities of the physical world.

1. Yuyong Technology: The entire capital circle is paying attention to Lin Junyang

Yuyong Technology was founded in May 2026 in Shanghai by Lin Jinyang, former head of Alibaba’s Qwen large model team, and is positioned as an AI lab focused on world models and embodied intelligence.

Lin Junyang, born in 1993, earned his bachelor’s degree from the Department of English at the University of International Relations and his master’s degree in Linguistics and Applied Linguistics from Peking University. He joined Alibaba DAMO Academy in 2019 and was promoted four levels within six years, becoming Alibaba’s youngest P10 technical expert at age 32.

During his time at Alibaba, he led the development of the Qwen large model, driving the number of derivative models of the Qwen series on Hugging Face, the world’s largest AI open-source community, to exceed 200,000 and total downloads to surpass 1 billion, securing the top position globally among open-source large models. According to Stanford University’s 2025 Artificial Intelligence Index Report, the performance gap between top AI models in China and the U.S. has narrowed to 0.3%, with Qwen’s core models ranking third globally in contribution.

In early March 2026, Lin Junyang officially left Alibaba.

Yuyong Technology completed its seed funding round just two months after its founding, with a post-money valuation of approximately $2 billion. Investors include Sequoia Capital China, Tencent Investment, and Gaorong Capital.

On July 13, Yuyong Technology completed its angel round industrial and commercial changes, adding Tencent’s Shanghai Qishan Investment Co., Ltd., HongShan Growth VII Holdco B Ltd. (Sequoia China), and Xiamen Yaheng Venture Investment Fund as new shareholders, with registered capital increasing from RMB 250,000 to RMB 1.25 million.

Given that the first-round valuation reached unicorn status, the capital market has expressed very high expectations for Lin Junyang’s team’s world model approach.

The company is currently focused on embodied intelligence and world model data processing, as well as foundational model development, and does not plan to engage in raw data collection or ontology-related businesses. It is actively seeking technical collaborations with local Shanghai institutions such as Shanghai Chaizhi Academy and Fudan University.

Lin Junyang's departure to start his own venture has drawn attention in the industry not only because of his status as a senior executive leaving a major company, but also due to his choice of direction—not building another general-purpose large model, but focusing instead on world models and embodied intelligence. As the driving force behind Qwen’s rise to become the number one open-source AI model globally, he clearly has his own vision for where AI should go next.

2. Physis (Inverse Matrix Technology): Spun out from the Beijing Academy of Artificial Intelligence, led by Gen Z founders

Inverse Matrix Technology may be the most unique among these 23 companies.

The company was co-founded in February 2026 by Chen Boyuan, an undergraduate at Yuanpei College, Peking University, and Ji Jiaming, a Ph.D. candidate at Peking University, who are fellow students under the same advisor.

Chen Boyuan, 22, received the best paper award at ACL and a spotlight paper distinction at NeurIPS during his undergraduate studies, and was named Peking University Student of the Year 2025, receiving the university’s highest honor, the May Fourth Medal.

Ji Jiaming is one of the few young researchers to have received doctoral scholarships from Apple, Tencent, and Ant Group. He has over 6,000 citations on Google Scholar and was also named Peking University’s Person of the Year. The team is composed almost entirely of Gen Z members born in the 1990s and 2000s, with core members having published over 80 papers at top international AI conferences including NeurIPS, ICML, ICLR, and ACL. Their open-source projects have garnered over 20,000 stars on GitHub and have been downloaded more than five million times.

Inverse Matrix is a team incubated by the Beijing Academy of Artificial Intelligence. The English name, Physis, derives from the Greek root for "nature," symbolizing a return to the fundamental principles of physics to understand how the world operates.

The Zhiyuan Institute has recently established an Innovation Center for Behavioral World Models, led by Chen Boyuan, who reports directly to Zhiyuan Institute President Wang Zhongyuan.

On the technical roadmap, the inverse matrix deeply integrates world foundation models with reinforcement learning, extending them into a physical latent space where action primitives natively enter the latent space.

On June 12, at the opening ceremony of the main forum of the 8th Beijing Zhixuan Conference, Inverse Matrix officially launched the world’s first general-purpose world foundation model, “WuJie·Physis-v0.1.” Centered on predicting the next physical state, this model addresses industry-wide challenges such as current models’ lack of understanding of real physical laws, low reliability of simulation outcomes, and missing long-term temporal memory. It establishes a technical closed loop through “compression-action causality-reinforcement learning validation-generalization,” and has already been deployed in real-world physical scenarios including embodied intelligence, physical simulation, interactive worlds, and scientific prediction.

Chen Boyuan proposed a W0 to W5 grading system for world models: W0-W1 address video generation and interactive worlds; W2's core feature is "true understanding of physics"; W3 solves cross-domain capabilities; W4 represents true generalization.

From a funding perspective, Inverse Matrix Technology is one of the companies with the most active capital movements in this group:

In March, the company completed a multi-million-dollar seed round led by Yanyuan Venture Capital, Hillhouse Capital, and Lingchu Intelligent Investment. In June, it raised over $100 million in a seed++ round, with new investors including BAI Capital, Wuyuan Capital, Hapu Capital, Yanyuan Venture Capital, Matrix Partners China, Ant Group, Zhongding Capital, and Hillhouse Capital. Gaohu Capital served as the exclusive financial advisor.

The emergence of Ant Group signifies that internet giants' world model strategies have expanded from purely in-house development to external investments. Invert Matrix Technology has been labeled a "Thousand-Mile Horse Project" by IT Juzi, and according to media reports, its valuation has exceeded RMB 5 billion, making it one of the highest-valued AI companies founded by Generation Z.

The 22-year-old founder, who started the company while still an undergraduate—Inverse Matrix Tech’s very existence is a fascinating footnote in the world models赛道.

These young entrepreneurs share a common trait: rather than gaining experience in the industry before starting their ventures, they began commercializing their work directly at the forefront of academia.

The technological cutting-edge of world models means that this field places far greater weight on "academic depth" than on "industry experience," and Inverse Matrix Technology is an extreme embodiment of this logic.

Other companies in the same group

PowerAI was founded in January 2026 in Beijing by Professor Feng Shuo of the Department of Automation at Tsinghua University. Professor Feng returned to China in 2022 to join Tsinghua, and within three years was promoted from assistant professor to associate professor and doctoral supervisor by special appointment. He authored a cover article in Nature as first author and was selected by MIT Technology Review China’s “35 Under 35” for innovation in technology. The company’s technical approach is “Next State Prediction”—using generative methods to predict the next state of the physical world. Its flagship product is named DenseWorld, and Feng’s vision is to build “the Claude of Physical AI.” In June, the company completed its angel round of funding, with investors including Qianyuan Kunyi, StarLink Capital (Z Fund), Spring Capital, and Tsinghua Ventures. The participation of Spring Capital, a major private equity firm, in an early-stage investment is noteworthy.

Lan Zhi Tan was founded in May 2026 in Beijing by Tian Keyu, winner of the Best Paper Award at NeurIPS. Tian’s NeurIPS 2024 Best Paper, VAR (Visual Autoregressive Model), has garnered significant attention in the field of video generation. The company was incubated and invested in by Yuan Capital and completed a multi-million-dollar seed round in July, with a valuation of approximately $200 million.

Shi Ke Xing was founded in July 2026 in Haidian, Beijing, focusing on the development of foundational models for Physical AI, and has since completed its seed funding round. The founding team’s background and product details have not been disclosed.

Add a timeline and depth to the world model.

This group of companies enters the world model space through 3D spatial reconstruction or 4D spatiotemporal modeling, with the core idea of creating digital twins of the physical world to provide training and simulation environments for AI.

Their technical foundation often involves 3D reconstruction techniques such as Gaussian Splatting and Neural Radiance Fields (NeRF), with additional time dimensions and physical rules applied on top.

Yuanhao Dynamics (PrimalVerse): Originating from the Institute for Intelligent Industry at Tsinghua University

Yuanhao Dynamics was founded in January 2026 in Beijing by Professor Zhao Hao, an assistant professor at the Institute for Artificial Intelligence and Industry (AIR) of Tsinghua University.

The core team comes from top universities such as Tsinghua University, Peking University, Shanghai Jiao Tong University, and HKUST, with long-term expertise in 3D/4D reconstruction, neural rendering, physics simulation, and robotics technology across the four foundational tech stacks: reconstruction, simulation, rendering, and robotics.

Recipient of multiple Best Paper-level achievements: SlimmeRF in real-time neural rendering won Best Paper at 3DV 2024; MARS in controllable simulation was Runner-up for Best Paper at CICAI 2023 (the world’s first open-source high-fidelity autonomous driving simulator); Dexora in robotics was a Best Paper Finalist at ICRA 2026 (the world’s first open-source high-degree-of-freedom dexterous hand VLA model); TRELLIS.2 in 3D reconstruction won Best Student Paper at CVPR 2026 (the world’s first single-image generator of 8K 3D models).

Yuanhao Dynamics introduces the concept of the "Fifth-Generation Multimodal 4D World Model," unifying 3D space and time with the native 4DWorldToken at its core.

In terms of commercialization, Yuanhao Power has moved ahead of its peers:

Collaborating with Jingfeng Medical on joint development of a world model specialized for medical scenarios, partnering with Geek+ to co-develop the Gravity4D world model for warehouse operations, and collaborating with Great Wall Motors on 3DGS training technology.

The company completed its seed round in July, raising hundreds of millions of RMB, with investors including Zhuoyuan Asia, Qifu Capital, Legend Star, Ginkgo Capital, and Jingfeng Medical.

Jingfeng Medical appears on the investor list as a medical device company, suggesting that Yuanhao Power’s 4D world model has made substantial progress in medical surgery simulation scenarios. Yuanhao Power is marked as a Qianlima project.

Other companies in the same group

Delta Intelligence was founded in January 2026 in Haidian District, Beijing, and is an embodied intelligence company incubated by the Beijing General Artificial Intelligence Institute (BIGAI), established by Professor Zhu Songchun.

The core product direction is a native humanoid foundation model (HFM), integrated with our proprietary native 3D world engine, aiming to achieve general-purpose whole-body coordinated operation.

The company’s fundraising pace has been highly intensive: in April, it closed a seed round of nearly RMB 100 million, with investors including Zhiyuan Robotics, Hillhouse Capital, Lujup Robotics, and Xinghaitu Investment; by the end of June, it sequentially closed multiple angel rounds, expanding its investor base to include Yuanhe Capital, Huazhang Capital, Huagai Capital, Fosun锐正 Capital, Lenovo Capital, Plum Ventures, and Haiwang Capital.

Three established robotics companies—Agibot, Looxid, and StarMap—simultaneously appear as seed-round investors, creating a rare scenario where portfolio companies invest in each other, signaling that upstream and downstream players in the robotics supply chain are accelerating the integration of world model technologies through capital ties. Delta Intelligence has been designated as a Qianlima Project.

Zhongke Yukun was founded in January 2026 in Beijing by Dr. Wu Yihong, a researcher at the Institute of Automation, Chinese Academy of Sciences, with Dr. Wei Hao serving as CEO. The company focuses on integrated 4D Gaussian reconstruction and decision-action world models, and completed a multi-million RMB seed round in June, exclusively invested by Shunxi Fund.

Help the world model understand "why"

This group of companies approaches world models through causal reasoning, with the core belief that a true world model must not only predict “what will happen next” but also understand “why it happens.”

They believe that generative models based on statistical correlations have fundamental limitations, and only by incorporating causal reasoning can AI achieve true physical understanding.

Aether AI / Original Wisdom: Startup founded by an Assistant Professor at UCSD and an Apple Scholar

Aether AI and Yuan Shi Zhi Zhi were both founded by the same individual—Professor Biwei Huang, an assistant professor at UCSD, an Apple Scholar, and a student of Clark Glymour, the founding figure of causal discovery.

The company was registered in March 2026 under the name Yuan Shi Zhi Zhi, launched operations under the Aether AI brand in June, and completed a $20 million funding round led by Matrix Partners, with academic advisors including Turing Award laureate Judea Pearl and other leading scholars from three generations of causal AI research.

Yuan Shi Zhi Zhi's technology is primarily applied in general robotics, AI agents, autonomous driving, and industrial automation, aiming to enhance the autonomous learning and adaptive capabilities of embodied intelligent systems through causal reasoning algorithms.

The world model empowers the robot

This group of companies uses humanoid robots or embodied intelligence as their core platform, treating world models as the robots' "cognitive brain" or "training engine."

Their common logic is: for a robot to act autonomously in the real physical world, it must possess the ability to understand, predict, and simulate its environment—and world models serve as the foundational infrastructure enabling this capability.

PokeBot: Founded by the Chief Scientist of StarMap

Breakout Robotics was founded in March 2026 in Haidian, Beijing, by Assistant Professor and Ph.D. supervisor Xu Huazhe from the Interdisciplinary Information Institute at Tsinghua University.

Xu Huazhe is one of the "Four Returnees from Berkeley," holding a Ph.D. from the University of California, Berkeley. During his postdoctoral research at Stanford’s Vision and Learning Lab (SVL), he proposed the 3D Diffusion Policy method—replacing 2D visual inputs with 3D visual information in Diffusion Policy to enhance policy generalization. This approach has become one of the widely adopted architectures in 3D visual imitation learning.

Over the past four years in China, Xu Huazhe has published more than a hundred high-quality papers, with results appearing in top international journals and conferences such as Science Robotics, T-RO, IJRR, NeurIPS, ICLR, CoRL, and ICRA. He has received the CoRL'23 Best System Paper Award and nominations for Best Paper at RSS'25 and ICRA'26. He is one of the few roboticists in China with end-to-end expertise spanning perception, decision-making, control, foundational models, and reinforcement learning.

He founded China’s first embodied intelligence lab focused on robotic manipulation, the TEA Lab, at Tsinghua University. In 2023, Xu Huazhe joined StarMap as a co-founder and chief scientist, helping transform the company into a leading name in China’s embodied intelligence field. At the peak of StarMap’s valuation, which surpassed 20 billion RMB and raised nearly 3 billion RMB in funding, he left the company at the end of 2025 and founded Breakout Robotics in March 2026, as home scenarios were not a primary focus for StarMap at the time.

The technical approach of Breakout Robotics differs fundamentally from industry norms: Xu Huazhe abandoned the VLA (Vision-Language-Action) foundation model framework and instead developed a world model that directly inputs and outputs "video-action," introducing a unique UAG architecture—replacing traditional cascaded pipelines with parallel pre-training and integrating reinforcement learning throughout both pre-training and deployment, achieving a fivefold improvement in training efficiency.

The first-generation embodied world model with 32 billion parameters has completed its initial training. In terms of product design, Breakthrough has opted for a wheeled base with dual arms rather than a humanoid bipedal form, directing all engineering resources toward fine manipulation capabilities of the arms.

Just three months after its founding, PoKe released a nine-minute video showing a robot independently cooking a dish of mapo tofu—from cutting the tofu and stir-frying the minced meat to adding seasonings and plating the final dish—executing dozens of interconnected steps with real-time error detection and correction. The video also demonstrated high-dexterity, long-sequence tasks such as folding clothes, threading zip ties, and tying sachets.

The pace of its funding was equally rapid:

In the month of April, the company completed a tens-of-millions-of-dollars angel round led by Yunqi Capital, with participation from Shunwei Capital, Honghui Fund, Xiaomi Strategic Investment, StarSea Map, Baidu Ventures, Innospace Angel Fund, Tsinghua Alumni Seed Fund, and Dongfang Jiafu. On August 3, it closed a nine-figure Pre-A round co-led by Shunwei Capital and Matrix Partners, with participation from Zhongding Capital, Jiukun Ventures, Junshan Capital, SEE Fund, Liepin Investment, and Yuannuo Capital. Existing investors—including Yunqi Capital, Xiaomi Strategic Investment, Honghui Fund, Innospace Angel Fund, and Dongfang Jiafu—all further increased their investments.

The Breakthrough Robot has been designated as the Qianlima Project, with a core team of approximately 20 members from Tsinghua University's Embodied Intelligence Lab. Xu Huazhe estimates that embodied intelligence will enter households in about 18 months—significantly faster than the industry's general expectation of 3 to 5 years.

Leaving a star company valued at $2 billion, Xu Huazhe chose a technological path radically different from the industry mainstream, avoiding the popular humanoid bipedal form—his consecutive "counter-consensus" decisions stem from a deep understanding of robotic operational scenarios. The mapo tofu he independently prepared in just nine minutes may be the most direct proof of this technical approach.

Kunlun Xing Robotics: Founded by executives from Huawei, Alibaba, and Li Auto

The founding team of Kunlun Robotics may have the most diverse industrial background among these 23 companies.

Founder and CEO Ren Geng has an extensive career spanning two industry giants, Huawei and Alibaba: He began his career at Huawei, where he became one of the youngest country general managers in Huawei’s overseas operations; in 2015, he joined Alibaba, serving successively as President of Alibaba Cloud China, President of Research and Development for Alibaba Cloud’s Video Cloud and Edge Cloud, and Chief of Staff of Alibaba Cloud, becoming one of Alibaba Group’s youngest vice presidents born in the 1980s; in late 2024, he assumed the role of President of New奥 Group, leading a large enterprise with over 50,000 employees.

Co-founder and CTO Lang Xianpeng was Ideal Auto’s “Employee No. 1,” having previously served as Senior Vice President and President of Autonomous Driving. He built Ideal’s autonomous driving R&D system from scratch, led the “Weicheng Project” to drive in-house development of Ideal’s autonomous driving technology, and successfully achieved mass production and deployment of highway NOA and urban NOA. He is one of the few industry leaders who have successfully taken AI-driven autonomous driving technology from concept to reality at scale.

Ren Geng excels in large-scale commercialization and supply chain management, while Lang Xianpeng specializes in world models and autonomous driving algorithms. This combination of “cloud-based commerce + autonomous driving algorithms” addresses the skill gaps present in most humanoid robotics startups today.

The company, with "physical intrinsic causality" as its core technology, adheres to a dual-wheel strategy of "ontology + brain," innovating and iterating on the VLA model by integrating physical causal modeling capabilities to build a dual-system robotic intelligence architecture centered on the Kunlun World Model (KWM).

Kunlun Xing has also established an industry-first counterfactual evaluation system that accurately distinguishes the true generalization capabilities of models using real physical outcomes, combined with "one brain, multiple forms" cross-ontology transfer technology and a software-defined hardware forward development model.

Kunlun Xing's fundraising journey has been a textbook example of rapid advancement:

In the month of its founding in March, it secured seed funding from Sequoia Capital China and Hillhouse Capital. In May, it received additional investments from China Star Venture, Huaye Tiancheng, and Hillhouse Capital. In June, it completed its Pre-A round, raising billions of RMB in total and achieving a post-money valuation exceeding $1 billion. In July, Kunlun Xing secured strategic investment from Xin Capital and Jianfa Emerging Investment, with all initial investors fully participating in subsequent rounds, and no investor exited throughout the process.

Kunlun Xing is the most capital-backed project in the 2026 world models赛道 and the fastest-funded unicorn. The company is headquartered in Yizhuang, Beijing, leveraging the local robotics industry ecosystem to accelerate technological iteration and mass production.

Other companies in the same group

Yao Song, the founder of Zhengxing Innovation, is a serial entrepreneur, and this is his third startup.

In 2016, 24-year-old Yao Song, after graduating from the Department of Electronic Engineering at Tsinghua University, co-founded Deephi Technology, an AI chip company, with his undergraduate advisor Professor Wang Yu and Stanford University’s Dr. Han Song. Two years later, the company was acquired by Xilinx for $300 million, becoming China’s first AI chip startup to achieve an exit.

In 2020, Yao Song co-founded Eastern Space, a commercial aerospace company, and set multiple records, including developing the world's largest solid-fuel launch vehicle.

At Zhengxing Innovation, Yao Song’s partner is Yu Chao, assistant professor at Tsinghua Shenzhen International Graduate School; both are alumni of the same program, and Yu Chao led the development of the multi-agent reinforcement learning algorithm MAPPO, which has become a benchmark method in the field.

CP Group is also one of the co-founders of Zhengxing Innovation, providing the company with a clear pathway for deployment in consumer retail scenarios. The company’s core technology is the World Action Model (WAM), which leverages reinforcement learning to enable large-scale deployment of robots in real-world commercial and industrial environments, with each action planning cycle taking only about 187 milliseconds.

In June, the company completed its angel round, raising nearly $100 million from investors Sinotruk Technology, Charoen Pokphand Group, and iHealth Labs—three companies from traditional industries in electronics manufacturing, agriculture and food, and medical devices, respectively, rather than traditional VCs—suggesting that Zhengxing Innovation’s commercialization path leans more toward industrial application. The company has been designated as a Qianlima Project.

Li Yiming, founder of Liqing Intelligence, was born in 1997, earned his Ph.D. from New York University in 2025, and is currently an Assistant Professor and Ph.D. supervisor at Tsinghua University's School of Artificial Intelligence, becoming a Ph.D. supervisor before the age of 30.

During his PhD at New York University, Li Yiming published over 20 papers at top conferences such as CVPR and NeurIPS, and was awarded the NVIDIA Scholarship in 2024—awarded to only 10 recipients worldwide each year.

Li Qing chose a path with extremely high technical barriers: fully in-house development of data acquisition devices, data pipelines, model architectures, and physics engines. Through proprietary data acquisition devices, they reduced the cost per unit from the dollar range to the renminbi range, increasing data volume from the industry-standard hundreds of thousands of hours to millions or even tens of millions of hours.

Leveraging the Tsinghua platform, the company has assembled over fifty top talents from Tsinghua, Peking University, Zhejiang University, UC Berkeley, and the University of Southern California, with an average team age of just 22 to 23. In June, the company completed its seed round with hundreds of millions of RMB, backed by an exceptionally prestigious group of investors: SEE Fund, Century Golden Resources, Fengrui Capital, Starlink Capital (Z Fund), AgiRobot, Tsinghua Alumni Seed Fund, Sequoia Capital China, Shunwei Capital, Hillhouse Capital, and Lingxin Qiaoshou.

The strategic investments in Zhìyuán Robotics and Lingxin Qiaoshou indicate that the full-stack approach is gaining collaborative support from upstream and downstream sectors of the industry. The company has been designated as a Qianlima Project.

Mogain Technology was founded in Shanghai in May 2026 by a team from HKUST. Its flagship product is MoSkin, a full-body flexible tactile system based on electromagnetic metamechanics, enabling robots with physical contact perception. In July, the company completed a multi-million-dollar seed round led by Agibot, Sequoia Capital China, and Hillhouse Capital.

Dawn Intelligence was founded in June 2026 in Shanghai by Yang Shuo, a 26-year-old tenured professor at Harbin Institute of Technology (Shenzhen). Yang Shuo is a recipient of the Google PhD Fellowship and a finalist for the ICLR Best Paper Award. The company’s technical approach centers on the TouchWorld tactile world model, aiming to build a complete capability chain for full-body mobility and dexterous manipulation in humanoid robots. In June, it completed a seven-figure RMB seed round led by Shunxi Capital.

Yiseng Brain was founded in Shenzhen in June 2026 by Han Yatong, a Ph.D. jointly trained by Harbin Engineering University and Stanford University. Its product, IS-Brain Cognitive World Model, integrates active inference mechanisms with hippocampal memory architecture, organizing perception, memory, reasoning, planning, decision-making, and action feedback into a continuously operating cognitive loop. Skills learned and anomalies encountered by one robot can be shared among all robots. On July 24, it completed a tens-of-millions-RMB seed round led by Haiyuan Capital.

World models achieve breakthroughs in specialized domains through narrow gates

This group of companies applies world models to specific vertical domains, including scientific research and development, life sciences, drug discovery, spatial visualization, data collection, and logistics and warehousing.

Their technical approach does not aim for generality, but rather validates the core capabilities of world models in specific scenarios—understanding, predicting, and simulating the physical world.

Quantum Power: Over 100 Million Yuan in Seed Funding

Quantum Dynamics was founded in March 2026 in Hangzhou by Li Qiang, former CTO of Cainiao Group and CTO of Alibaba International Digital Commerce.

Li Qiang has 17 years of experience within the Alibaba ecosystem, having led the end-to-end establishment and operations of systems across four key areas: Taobao and Tmall, Cainiao, international business, and autonomous vehicles. He has managed thousands of R&D professionals globally and concurrently served as General Manager of Cainiao’s Autonomous Vehicles, overseeing the development and commercialization of autonomous driving technology.

During his leadership of Cainiao's technology, Li Qiang found that the previous generation of automation technology could only handle storage and movement of standard containers and boxes, but struggled with tasks requiring hand-eye coordination—such as weighing and labeling packages, flipping items for sorting, restocking shelves, picking and removing items from shelves, and boxing and packing—precisely the areas in the logistics network with the largest labor shortages and highest labor cost proportions.

Quantum Dynamics is positioned as a Physical AI platform company, with its core product being an embodied-native World-Action Model—a model capable of generating actions, predicting causal relationships in the world, and evaluating whether states align with business semantics, also known as a World Value Model.

Li Qiang revealed that the company will launch a world model with 3-4 billion parameters, and the first-generation robot is expected to be released before September 30.

The company’s team spans three dimensions: technology, product, and industry. The CTO is the head of online algorithms at a leading autonomous driving company. The Chief Scientist is an assistant professor at a top domestic institution, specializing in foundational models for embodied intelligence. The head of hardware comes from the humanoid robotics division of a publicly listed company, with extensive academic and practical experience in robotic arms and dexterous hands at the bachelor’s, master’s, and doctoral levels. The commercialization lead graduated from the Physics Department at the University of Maryland and has served as an algorithms expert at Goldman Sachs, Amazon, and Meta.

The company has established industry-academia-research collaborations with universities such as Zhejiang University, Nanyang Technological University, Shanghai Chuangzhi Institute, and the Hong Kong University of Science and Technology.

Clear commercialization path: In 2026, focus initially on the three core roles with the highest labor intensity—restocking, picking, and packing—in B2C and small B2B warehouses; after 2027, expand to express delivery outlets, ultimately progressing along the path of “logistics → logistics-like scenarios such as industrial production lines, retail restocking, pharmaceutical sorting → business services → public services such as elderly care and household support.”

Completed a seed round exceeding RMB 100 million on July 31, co-invested by Yunqi Capital and SenseTime Guoxiang Capital.

What sets Quantum Power apart is the founder’s deep understanding of real-world scenarios. With 17 years at Alibaba, Li Qiang possesses on-the-ground insights into which logistics processes are poorly suited for automation, which roles face the greatest workforce shortages, and which scenarios are best suited for the initial deployment of embodied intelligence.

This problem-driven entrepreneurial approach gives Quantum Dynamics a clearer commercialization path than teams with purely technical backgrounds.

Other companies in the same group

Field Sampling was founded in 2026 in Shanghai by Dr. Peng Xiangda from the Dalian Institute of Chemical Physics, Chinese Academy of Sciences, introducing world models into the field of molecular dynamics simulation. During his postdoctoral research at the University of Chicago, Peng Xiangda collaborated with John Jumper, one of the leads of AlphaFold and a Nobel Prize laureate in Chemistry, to develop a series of machine learning molecular force fields.

The company's new model integrates classical force fields with machine learning, achieving a fivefold speed improvement at similar accuracy. Within one month of its founding, it secured two R&D contracts from innovative pharmaceutical companies and is one of the featured projects at Qihai Innovation's 2026 Spring Roadshow.

Qingxi Zhiguan was founded in Beijing in May 2026 by a team with backgrounds from Tsinghua University and Stanford University. The company introduced the concept of a "Scientific World Model," building an end-to-end, fully integrated AI self-evolving architecture for scientific discovery and industrial R&D in fields such as materials, chemicals, energy, and semiconductors. In July, it completed a multi-million-yuan seed round led exclusively by WuYueFeng Tech.

Philo AI was founded in Beijing in 2026 by Zhang Jia Sheng, a Ph.D. from Tsinghua University's Institute for Interdisciplinary Information Sciences and a former "Genius Youth" at Huawei. He introduced the concept of the "World Life Model," aiming to enable video characters to engage in real-time, two-way interaction with users. Zhang previously developed the Avenger video model, which ranked second globally. In June, the company completed a seven-figure USD seed round led exclusively by Xiangfeng Capital.

Linglong Vision, founded in 2026 in Shanghai by Dr. Ma Weijie of Fudan University, leverages his naked-eye 3D display technology published in Nature to serve as a seamless visual medium for physical intelligence, creating an immersive,眩晕-free, ultra-wide-angle interaction container that requires no wearable devices. It is one of the projects featured in the 2026 Spring Demo Day of Jiqi Chuangtan.

Si Ai Technology was founded in 2026 in Beijing by Zhang Zhening, a former Tesla and Amazon Go engineer. Leveraging a crowdsourced network of food delivery riders to collect in-door multimodal data and developing proprietary AI data collection hardware, the company has accumulated nearly 100 TB of EgoVideo data, covering high-precision maps of key commercial areas in Beijing and Shanghai. It was one of the projects featured in the 2026 Spring Demo Day of Jiqi Chuangtan.

01 Capital made a significant bet at the seed stage and is still accelerating.

In 2026, the funding characteristics of these companies were "early" and "large."

Yuyong Technology raised hundreds of millions of USD in its seed round, Inverse Matrix Technology raised over one hundred million USD in its seed round, Kunlunxing Robotics raised several billion RMB in its Pre-A round, and Poke Robot raised hundreds of millions of USD in its Pre-A round—these amounts typically correspond to Series B or even Series C levels in traditional startup funding cycles.

Notably, over the past ten days, the pace of funding has not slowed—it has accelerated: Yiseng Brain completed its seed round on July 24, Quantum Power closed a seed round exceeding RMB 100 million on July 31, and PoKe Robot secured a Pre-A round in the hundreds of millions of USD on August 3.

Capital markets are conducting price discovery in the world models赛道 at an unprecedented speed and scale, likely driven by the core logic that world models are widely regarded as the next AI paradigm following large language models—and the window for paradigm-shifting opportunities is extremely short, making the early entry cost high but the cost of missing out even higher.

02 The founders have an extremely diverse background.

The founders of these 23 companies have backgrounds of remarkable diversity, rare in the history of AI entrepreneurship in China.

Academic stars: First author of a Nature cover paper (Feng Shuo), lead of Alibaba’s Qwen (Lin Jinyang), Peking University Person of the Year (Chen Boyuan), Tsinghua University’s youngest PhD supervisor (Li Yiming), and founder of Tsinghua’s Interdisciplinary Information Sciences TEA Lab (Xu Huazhe).

Backgrounds from leading companies: Ren Geng, former President of Alibaba Cloud China; Lang Xianpeng, President of Ideal Autonomous Driving; Li Qiang, former CTO of Cainiao.

In terms of serial entrepreneurs: Yao Song, who has worked on AI chips and commercial rockets, and Xu Huazhe, who left a star company valued at 20 billion.

The trend toward younger leadership is particularly notable: Chen Boyuan is 22, Feng Shuo is under 35, Li Yiming is under 30, and Yang Shuo is 26, with several teams averaging between 22 and 28 years old. A significant proportion of these founders are Gen Z or even Gen Alpha, and they are becoming the main force in the world models赛道.

03 Capital in the industry chain proactively builds an ecosystem

Agibot appears among the investors of Delta Intelligence, Liqing Intelligence, and Mogan Technology. Looj Robotics and StarMap appear in Delta Intelligence’s seed round. Lingchu Intelligence appears in Nijuzhen Technology’s seed round. StarMap also appears in Poko Robotics’ angel round. Xiaomi Strategic Investment participates in Poko Robotics’ angel and Pre-A rounds.

These established robotics companies and smartphone manufacturers are building embodied intelligence ecosystems centered around themselves by investing in startups specializing in world models and perception technologies.

For the invested company, securing investment from industry players means not only capital but also access to real-world data and deployment opportunities—both of which are critical for training and validating world models.

04 Technical pathways have clearly diverged

23 companies do not have a unified understanding of "world models."

The generalist base camp seeks modeling capabilities across diverse physical-world scenarios; the 4D spatial intelligence camp extends 3D reconstruction by adding a time dimension; the causal reasoning camp emphasizes understanding causal relationships in the physical world; the embodied intelligence camp treats world models as the cognitive engine for robots; and the vertical scenario camp validates the core capabilities of world models in specific domains such as scientific discovery, molecular simulation, spatial display, and logistics and warehousing.

This divergence indicates that world models are still in the conceptual exploration phase, and no single approach has been proven as the only correct path.

It is worth noting that Breakthrough Robotics and Kunlun Mobility, both belonging to the embodied intelligence camp, differ in their technical approaches:

Pohai has completely abandoned the VLA model in favor of world models, while Kunlun Xing has introduced physical causal modeling capabilities on top of mainstream international VLA models.

For investors, diversifying bets across multiple paths is a sound strategy—Jingjin Ventures has invested in both NiJuzhen Technology and Aether AI, Shunwei Capital has invested in both PoKe Robotics and LiQing Intelligence, and Zhongding Capital appears on the investor lists of NiJuzhen Technology, Kunlun Xing, and PoKe Robotics.

05 Deep Integration of Academia and Industry

Several of these companies were directly incubated by academic institutions:

NiJuzhen Technology was incubated by the Beijing Academy of Artificial Intelligence; Detta Intelligence was incubated by BIGAI; Zhongke Yukun was incubated by the Institute of Automation, Chinese Academy of Sciences; Yuanhao Power originated from Tsinghua University's AIR; Mijí Intelligence originated from Tsinghua University's Department of Automation; Liqing Intelligence originated from Tsinghua University's School of Artificial Intelligence; Posun Intelligence originated from Harbin Institute of Technology (Shenzhen); and Yixin Brain Team comes from Stanford University, the University of Science and Technology of China, and The Chinese University of Hong Kong, Shenzhen.

Founders typically maintain their academic or research positions while launching startups, creating a "professor + CEO" dual-track model.

Chen Boyuan also serves as the head of the Behavioral World Models Innovation Center at the Beijing Academy of Artificial Intelligence, reporting directly to President Wang Zhongyuan, in his capacity as founder.

This deep integration of academia and industry is particularly essential in the world models赛道—core technical challenges in world models remain at the frontier of research, and the depth of academic accumulation directly determines the ceiling of technological development.

06 Talent is rapidly leaving major tech companies

Lin Jinyang from Yuyong Technology comes from Alibaba’s Tongyi Qianwen; Li Qiang from Quantum Dynamics comes from the CTO of Cainiao; Xu Huazhe from Breakthrough Robotics comes from the Chief Scientist of StarMap. All these founders left core positions at established companies to start their own ventures.

More notably, Tencent has built its investment landscape in the world model space through investments in projects such as Pragmatic Tech and PoKe Robot; Ant Group has directly invested in Inverse Matrix Tech; Xiaomi’s strategic investment arm appears among the investors of PoKe Robot; and SenseTime has invested in Quantum Dynamics through Guoxiang Capital.

Major tech companies' world model strategies have expanded from purely in-house development to systematic investment, with a dual approach of "in-house development + investment" becoming the standard strategy.

07 Beijing Haidian remains the absolute center, but a clear trend toward multi-point dispersion is evident.

Of the 23 companies, 15 are located in Beijing, 6 in Shanghai, and one each in Hangzhou and Shenzhen.

The concentration in Haidian District is extremely high, due to the presence of core academic institutions such as Tsinghua University, the Chinese Academy of Sciences, BIGAI, and the Beijing Academy of Artificial Intelligence, along with a large number of AI professionals.

However, the new spreading trend is worth noting:

Hangzhou has seen the emergence of Quantum Dynamics (founded by former Cainiao talent), Shenzhen has witnessed the rise of Yixin Intelligence (backed by CUHK-Shenzhen and Shenzhen’s robotics industry ecosystem), and Beijing’s Yizhuang district has attracted Kunlun Mobility (hosting over 300 companies in its robotics industry cluster), indicating that world model talent is dispersing from a single hub to multiple nodes.

The 23 world model companies founded in the first seven months of 2026 signal a clear shift: world models have moved from conceptual discussion into the phase of engineering implementation.

The founders' backgrounds, funding sizes, and technology choices of these companies indicate that the industry is no longer content with debating "what is a world model," but is instead voting with their feet—defining what a world model should be through code and capital.

In the first seven months of 2026, China’s world model startup landscape delivered: 23 newly founded companies, 5 distinct technical pathways, 2 unicorns, and 4 racehorses.

The breakout robot achieved its nine-digit Pre-A round in less than five months from inception, and Kunlun Xing became a unicorn in just 90 days from registration—both setting new speed records for AI startups in China.

It is still unclear how many of these companies will ultimately succeed.

What is clear is that the competitive landscape in the world models space is shifting from “who is talking” to “who is doing,” and this new generation of players has already seated themselves at the table for 2026.

This article is from the WeChat public account "IT Juzi" (ID: itjuzi521), authored by Judy.