Liang Wenfeng has made another move.
From Unitree to Winverse AI, Liang Wenfeng's AGI roadmap is accelerating
On October 12, an exclusive report from Bloomberg instantly pushed a company named Winverse AI into the public spotlight.
According to people familiar with the matter, DeepSeek has agreed to lead its first round of financing, with the seed round scale reaching tens of millions of U.S. dollars and a post-money valuation of about 400 million U.S. dollars. Longzhu Investment under Meituan has also made an investment commitment. The news spread over and over again in investors' WeChat Moments and WeChat groups. The first question people cared about was not what Winverse actually does, but why Liang Wenfeng chose to make a move.
The founder at the helm of Winverse is Hu Han. He won the Marr Prize, one of the highest awards in the field of computer vision, and left Tencent Hunyuan this summer. Bloomberg stated that when he left, Tencent was restructuring its artificial intelligence research team to catch up with rivals such as DeepSeek and Moonshot AI.
The connection between the two sides is quite dramatic. Tencent wanted to catch up with DeepSeek, so it adjusted its team, and Hu Han left the company to start his own business; in turn, the person who wrote the first check for his new company was exactly DeepSeek. And Tencent was also the external investor that contributed the most in DeepSeek's first financing. After taking a big circle, the three parties returned to their original positions.
Two months earlier, DeepSeek's name appeared on the strategic placement list of Unitree Technology, and at that time it "rode along" with a soon-to-be-listed star company. This time, it directly took the position of the lead investor of the seed round. Reports say Liang Wenfeng once regarded embodied intelligence as the last stop on the road to AGI.
Marr Prize Winner, Former Tencent Researcher
Hu Han has an impressive resume in the field of artificial intelligence. He obtained his doctorate from the Department of Automation, Tsinghua University. In 2012, he conducted visiting research at the GRASP Lab of the University of Pennsylvania, then joined Baidu Institute of Deep Learning, and joined the Visual Computing Group of Microsoft Research Asia in December 2016, and had become the principal researcher of the group before leaving. For computer vision researchers of that era, Microsoft Research Asia was the best place to work. There, he and his team completed a work that has been cited many times later.
This paper is called Swin Transformer. Hu Han is the sole corresponding author and a co-first author. The paper won the Marr Prize, the best paper award at ICCV 2021. The conference received 6152 submissions that year, and only this paper was selected as the best paper.
A sentence Hu Han said in an interview back then seems very meaningful in retrospect. He mentioned that various sub-fields of artificial intelligence have long used their own basic networks, making it difficult to learn from each other on the same platform. The team has always wanted to contribute to unifying these research platforms.
Swin Transformer is designed to do exactly that. It brings Transformer, which has achieved great success in the language field, into visual tasks. Language and vision began to share a common foundation. The next step is to incorporate "action" into it, which is exactly the problem that embodied intelligence aims to solve.
At the beginning of 2025, Hu Han joined Tencent. He took the position left by Liu Wei, the former technical head of Tencent Hunyuan, whose departure was exposed in November 2024.
According to the report of Intelligent Emergence, after Hu Han joined Tencent, he first engaged in the research of large vision models. After the organizational adjustment, he entered the "Frontier" cutting-edge technology research group under the large language model department, responsible for multi-modal understanding, and reported to his superior Yao Shunyu. During this period, he also participated in the research and development of world models.
In July this year, he submitted his resignation, and Tian Yonglong, a former OpenAI researcher, took over his position. The background of this personnel change is the overall consolidation of Tencent's AI system. Industry insiders directly stated in the report that multi-modal understanding scenarios such as image recognition are difficult to charge directly. What users are really willing to pay for are office applications, and the competition behind them lies in reasoning, programming and Agent capabilities.
There is very little public information about what Winverse plans to do. According to the data sorted out by Dealroom, the AI model being developed by the company for its humanoid robot aims to enable the model to understand instructions issued by ordinary people and convert the instructions into actions that the robot needs to perform in unfamiliar environments.
Bloomberg classifies Hu Han as a group of Chinese AI researchers who are leaving large tech companies to build dedicated laboratories on their own. Many of these people are working on world models for autonomous driving, interactive videos or robots.
Tens of millions of dollars is a small sum for a company valued at tens of billions of dollars. But Bloomberg believes this will be one of the largest bets DeepSeek has placed on a startup. The amount of money is not large, but the signal significance is very important.
From Unitree to Winverse, Liang Wenfeng Makes Another Move
Liang Wenfeng's first step in the field of embodied intelligence fell on Unitree. On August 6, the IPO issue price of Unitree Technology was determined to be 150.8 yuan per share, corresponding to an issue market value of about 60.9 billion yuan. The operating entity of DeepSeek was included in the strategic placement list.
DeepSeek obtained 933,400 shares this time, with an amount of 141 million yuan. The lock-up period of this part of the shares is 36 months, which is the longest in the entire strategic placement list. Those who obtained the strategic placement quota together with it also include Kunlun Capital under PetroChina, Southern Power Grid Industrial Finance Holdings, Tianyi Capital under China Telecom, and Shanghai Qishan Investment under Tencent.
Calculated based on the total share capital of about 404 million shares after Unitree's issuance, DeepSeek's shareholding ratio is about 0.23%. Liang Wenfeng himself did not rest during the offline new share subscription. 155 products under High-Flyer Quant were allocated 195,000 shares, worth about 29.41 million yuan; 43 products under Jiuzhang Asset were allocated 63,200 shares, worth about 9.53 million yuan. The three parties obtained a total of about 1.19 million shares of placement, with a capital scale of nearly 180 million yuan.
What is worth noticing is the content beyond capital. Unitree disclosed during the roadshow that the two parties signed a "Memorandum of Strategic Cooperation". The cooperation content includes the research and development of general artificial intelligence, high-performance robots and large AI models.
According to the memorandum, Unitree and DeepSeek will give priority to each other under equal conditions to jointly carry out research on large models and embodied intelligence technologies; when DeepSeek conducts embodied-related businesses in the future, it will give priority to purchasing robots and related technical solutions from Unitree; when Unitree carries out large model adaptation and integrated development, it will give priority to choosing DeepSeek's model training services. DeepSeek will also provide Unitree with competitive support in model architecture and intelligent computing cluster construction. The media summed up this relationship very vividly: DeepSeek provides the "brain", Unitree provides the "body", and the two become each other's priority procurement targets. In fact, this is a two-way binding relationship.
According to reports from ifeng.com, Wang Xing once said that when it comes to making "bodies", Unitree ranks first in the world, but when it comes to making "brains", the competition becomes much more difficult. To a certain extent, this sentence can explain why Unitree chooses to cooperate with DeepSeek.
It is worth mentioning that on October 12, people familiar with the matter revealed that Liang Wenfeng has recently frequently contacted the embodied intelligence field, communicated with many entrepreneurs, and discussed the interaction direction of robots, large models and the real world. People familiar with the matter said that Liang Wenfeng is considering whether DeepSeek should further penetrate the embodied intelligence field, with options including forming his own team for R&D or entering through investment. DeepSeek has not responded to this.
DeepSeek has previously completed its first round of financing, with a financing amount of about 500 billion yuan, and Liang Wenfeng personally contributed 200 billion yuan; at present, news comes that DeepSeek's second round of financing is nearing the end, and the financing amount has increased from the previously planned 500 billion yuan to about 1 trillion yuan, with a valuation of about 5 trillion yuan.
Bloomberg said in its report on October 12 that the money from these financings of DeepSeek was originally used for model training and data center construction, but it also provides DeepSeek with a new option: to support some new startups and carry out joint research and development with them. In August, it was a strategic investor of a listed company, and by October, it became the lead investor of the seed round. There are only two months between the two steps.
Why Does Liang Wenfeng Make a Move Now
Liang Wenfeng's views on embodied intelligence have always been rather vague. Reports mention that Liang Wenfeng believes that the road to AGI starts from the chain of thought, followed by Agent, continuous learning, and self-iteration. Only after all these are completed will embodied intelligence come next. Another report mentions that Liang Wenfeng once said that embodiment is an unavoidable destination for DeepSeek. The meaning of the two statements is roughly the same: embodied intelligence is still a long way from DeepSeek, but it will come sooner or later.
Since the distance is still very far, why start spending money now? One statement is that DeepSeek does not participate in the competition, but uses investment to seize positions first.
Looking at DeepSeek's recruitment information, you can understand that its main force is always focused on other areas. A report in September said that DeepSeek is recruiting about 150 senior engineers, all positions are concentrated in the direction of server-side and Agent R&D, and there is no AI researcher position at all.
At the time of the first round of financing, DeepSeek's team was said to have only about 170 people. It is almost impossible for such a team to conquer language models, Agent, robot hardware and other fields. Handing over the "body" to Unitree, and handing over the exploration of the embodied brain to external teams like Hu Han's, while focusing on providing models and computing power, seems to be a more cost-effective deal.
It has become a trend for top researchers to leave large tech companies to start their own businesses. Lin Junyang left Alibaba, Hu Han left Tencent, and they did not choose to join another large tech company, but established their own laboratories. For DeepSeek, these people are both competitors and allies that can be won over.
It is getting more and more difficult to recruit top talents, but investment can make them "stay around". Tencent placed a bet on Lin Junyang, DeepSeek placed a bet on Hu Han. The talent competition among large companies is no longer as simple as sending offer letters, but requires signing investment agreements.
The capital flow of the entire industry also pushes DeepSeek forward. According to data, by the first half of 2026, the total financing amount of China's domestic embodied intelligence track has reached 437.66 billion yuan, of which companies focusing on "brains" received 222.53 billion yuan, accounting for more than half, while companies making robot ontologies only raised 56 billion yuan in the same period.
Major companies have already taken action around the ecological niche of "robot brain". Tencent Robotics X Lab has connected its self-developed Tairos embodied large model to the Unitree G1 robot. Its head Zhang Zhengyou publicly stated, "Tencent does not plan to make robot hardware, but only builds a platform to provide 'brains' for various robots."
Ant Lingbo under Ant Group has accessed more than 20 robot configurations from 17 manufacturers. In June 2026, Huawei Cloud jointly launched an ecological cooperation plan with more than 20 model manufacturers including Zhipu AI, DeepSeek and MiniMax. The story abroad unfolded earlier than in China. In 2024, Figure AI cooperated with OpenAI, but due to the limitations of large language models, this cooperation only lasted for more than a year. After that, Figure turned to independent R&D, and the company's valuation rose all the way to 39 billion U.S. dollars. Both large model companies and robot companies are looking for their ideal partners. If DeepSeek does not appear in this network, after the technical routes converge, it will only have one way out, which is selling APIs.
There are also data-related issues. The industry generally believes that the main bottleneck of embodied intelligence is not the problem of model parameters, but the problem of data from real physical world interactions. With Unitree's robots and Winverse's models, DeepSeek has mastered the two parts that make data operate.
However, skeptical voices always exist. There is a controversy over the foam in the embodied intelligence track. The fundamentals of many projects have not been significantly improved, but their valuations have doubled in a short period of time. Therefore, most investors choose the strategy of "prefer to invest at a high price rather than miss the opportunity".
Some investors also worry that when technical executives start their own businesses, they tend to face acclimatization on the road to commercialization. Hu Han's advantages lie in vision and multi-modal, while his team still needs to improve the capabilities in hardware, control and scenario delivery required by robots. Zhang Wei, founder of Exyn Dynamics, made an analogy: a pure large model is like Hawking lying in bed, very smart but unable to move. A CTO of an embodied intelligence company also said that if the technical paradigm is subverted, for example, shifting from VLA to world models, deep binding will become a burden.
Five years ago, Hu Han hoped to put all sub-fields of artificial intelligence on the same platform. Now he wants to integrate language, vision and action into the same robot. This transaction has not been finally confirmed, and Winverse has not launched any products yet. It may take several years to see whether Liang Wenfeng's move is a strategic first move or a casual chess piece.
This article is from the WeChat official account "Thecapital" (ID: thecapital), author: Wang Tao, editor: Wu Ren, published with authorization from 36Kr.