Author: Zen, PANews
Founded only a few months ago and without having publicly launched any product, a startup named Winverse AI recently secured a valuation of nearly $400 million. The force behind this push is none other than AI giant DeepSeek, which has been making frequent moves recently.
On October 12, according to Bloomberg, DeepSeek has agreed to lead Winverse AI's seed funding round of tens of millions of dollars, with institutions such as Meituan's Long-Z Investments also having made investment commitments.
What the market is paying more attention to is DeepSeek's investment direction, which has gradually become apparent recently.
Over the past two years, this AI company founded by Liang Wenfeng has mainly attracted attention for the performance, training efficiency, and open-source strategy of its large language models. Now, DeepSeek has begun to look beyond language models, engaging with embodied intelligence-related companies through investment and technical cooperation, and its investment path toward extending intelligence into the physical world is gradually becoming clearer.
Why did Winverse AI secure a valuation of nearly $400 million?
Winverse AI, which has won DeepSeek's favor, was founded this year and is still operating in stealth mode. The outside world knows very little about its team size, technology R&D direction, or business model.
By contrast, the technical background of Winverse AI founder Han Hu is more noteworthy.
According to information disclosed on Han Hu's personal academic homepage, he received his bachelor's degree and doctorate from Tsinghua University in 2008 and 2014, respectively. He previously worked at Baidu's Institute of Deep Learning and Microsoft Research Asia, and joined Tencent's Hunyuan large model team in December 2023.
During his time at Tencent, Han Hu served as the technical lead for language large models, built a speech model team, and subsequently led the vision-language model (VLM) and embodied foundation model teams. In other words, before founding Winverse AI, he had already accumulated research experience in large models related to language, vision, and embodied intelligence.
One of Han Hu's best-known academic achievements is his participation in proposing the Swin Transformer vision model architecture, for which he served as corresponding author of the related paper. That paper won the Marr Prize for Best Paper at the 2021 International Conference on Computer Vision (ICCV).
Through designs such as shifted window attention, Swin Transformer enables the Transformer architecture to process visual information at different scales more efficiently, and it has been widely applied in computer vision tasks such as image classification, object detection, and semantic segmentation. These technologies are not directly equivalent to robot control models, but they are technically connected to the visual perception capabilities robots need to understand their external environment.
In addition, Han Hu also participated in research including Swin Transformer V2 and SimMIM, with work covering areas such as vision model scaling and pretraining. Judging from these experiences, Winverse AI's founding team has a strong background in foundational model research.
Regarding Winverse AI's specific business positioning, Bloomberg described it as a startup developing AI technology for robots. Phoenix Technology reported that, according to what it learned from multiple industry insiders, Winverse AI is not strictly an embodied intelligence company, but leans more toward underlying infrastructure such as model acceleration.
From Unitree to Winverse, DeepSeek accelerates its embodied intelligence layout
Winverse is not DeepSeek's first bet in the robotics field. In August this year, DeepSeek completed an investment of about 141 million yuan by participating in the strategic placement in Unitree Robotics' STAR Market IPO.
According to the issuance and allocation results disclosed by Unitree Robotics, Hangzhou DeepSeek Artificial Intelligence Basic Technology Research Co., Ltd., under DeepSeek, was allocated 933,399 shares, with an allocated amount of about 140.76 million yuan, accounting for 2.31% of the shares issued in this IPO, with a lock-up period of 36 months.
More noteworthy than the investment amount is that the two sides also established a clear strategic cooperation arrangement, under which they will combine their respective technical capabilities in AI models, mechanical engineering, motion control, and embodied intelligence to jointly explore AI model R&D for humanoid robots.
Under the relevant cooperation arrangement, when Unitree Robotics purchases model training services and technical solutions, it will give priority to DeepSeek; when DeepSeek purchases robots or carries out embodied intelligence applications, it will also give priority to Unitree Robotics.
This makes the relationship between the two sides more than just capital-level. DeepSeek's strengths in model training, inference, and algorithm optimization are somewhat complementary to Unitree Robotics' capabilities in robot bodies, motion control, and actual hardware deployment. For research institutions hoping to bring large models into the physical world, robot platforms may also provide perception and interaction data in real environments.
Compared with Unitree Robotics, Winverse AI is at an earlier stage of development. If this funding round is completed, DeepSeek's investment targets will cover both robot hardware manufacturers and startup teams with backgrounds in vision-language models and embodied model R&D.
In fact, DeepSeek founder Liang Wenfeng has already begun to pay more attention to this field.
According to an exclusive report by Phoenix Technology on October 12, Liang Wenfeng recently exchanged views with multiple embodied intelligence entrepreneurs, discussing technical directions such as robots, large models, and interaction with the real world.
People familiar with the matter revealed that DeepSeek is evaluating whether to further enter the embodied intelligence field, including by forming its own R&D team or investing in related companies.
From language models to the physical world, what does DeepSeek intend to do?
DeepSeek is trying to extend large model technology into the physical world, and the hardware platforms, real-world scenarios, and interaction data held by robotics companies can precisely support this exploration.
In the existing technology system, large language models mainly learn knowledge and reasoning abilities through data such as text and code. But controlling robots to complete tasks requires solving problems across more dimensions. Robots not only need to understand instructions, but also to identify the spatial positions of objects, plan motion paths, control grasping force, and adjust actions in real time according to changes in the actual environment.
These capabilities involve 3D spatial understanding, motion control, action data, and physical feedback, and are problems that existing language models find difficult to solve independently.
Previously, DeepSeek had already accumulated a certain technical foundation in the multimodal model field. As early as December 2024, it had already released the DeepSeek-VL2 vision-language model series, supporting tasks such as visual question answering, image understanding, and visual grounding. However, there is still an obvious technical gap between vision-language understanding and robot motion control.
Therefore, cooperating with robotics companies such as Unitree Robotics may provide DeepSeek with real hardware platforms, interaction data, and technical validation environments. Investing in technology startups such as Winverse AI may help it access R&D capabilities in vision models, embodied foundation models, or related underlying infrastructure.
In addition to technical needs, DeepSeek's rapidly expanding capital strength recently has provided another condition for its extension into the robotics field. After completing its first external funding round of about $7.4 billion in June this year, DeepSeek launched a second funding round in July, initially planning to raise 50 billion yuan with a target valuation of 500 billion yuan.
But investor subscription enthusiasm far exceeded expectations. Bloomberg cited people familiar with the matter as saying that by early October, investment commitments obtained in DeepSeek's new funding round had exceeded 80 billion yuan, and according to signed investment terms, the final funding size could be close to 100 billion yuan. Tencent and CATL are among the main contributors, and the financing is expected to be completed in October.
If the deal is ultimately completed smoothly, DeepSeek will obtain a massive amount of new capital. This can not only support its continued investment in computing infrastructure, model training, and talent recruitment, but also means it will have more ample financial conditions to participate in R&D in other AI technology fields through investment and cooperation.
In the past, DeepSeek's funding support mainly came from Liang Wenfeng's High-Flyer Quant, and with the introduction of large-scale external financing, it has begun to have the ability to support external R&D teams through capital investment, no longer having to rely entirely on internal teams to explore new technical directions. This is similar to the approach of large AI companies such as OpenAI and Google in expanding technology applications through investment, cooperation, and ecosystem building.
From the two investments in Unitree Robotics and Winverse AI, the areas DeepSeek is focusing on cover robot hardware manufacturing as well as technical capabilities potentially linked to model R&D.For an AI company that has just secured large-scale capital backing, investing in robotics-related enterprises may be both a technical choice to explore the next stage of model capabilities and a way to establish external partnerships in the field of physical AI, where R&D cycles are long and technological pathways have yet to converge.
In this sense, DeepSeek's robotics layout may have only just begun, and its gradually ample capital reserves will also give this exploratory path more possibilities. Its technological journey has officially moved from the purely digital world toward the infinitely possible physical reality.