Liang Wenfeng has invested in the author of Swin Transformer, and DeepSeek has started to make arrangements for its embodied brain layout.

What? DeepSeek is pouring money into robotics companies again?!

After investing about 141 million yuan in Unitree Robotics in August this year, Liang Wenfeng has now set his sights on a newly established robotics AI startup —

Winverse AI.

According to Bloomberg, DeepSeek plans to lead its multi-million-dollar seed round financing.

The post-investment valuation is about 400 million US dollars, equivalent to about 2.8 billion yuan.

Winverse AI was established just this year. Its founder Hu Han once led research on vision-language and embodied AI foundation models at Tencent.

It is worth noting that this is already the second time DeepSeek has invested in the robotics sector in just two months.

In August this year, DeepSeek just invested about 141 million yuan in Unitree Robotics, and the two sides also plan to carry out cooperation around embodied intelligence, robotics and large models.

Now, DeepSeek has shifted its investment focus to the AI model team behind robotics.

If the financing is completed, it will become one of the largest startup investments DeepSeek has made since its establishment.

Former Tencent Scientist Launches New Business, DeepSeek Leads Multi-Million-Dollar Investment

Let's first take a look at the company that DeepSeek is willing to invest in.

Winverse AI was founded in 2026. It is still in the early R&D stage and has not publicly disclosed its specific products and technical routes.

In contrast, its founder Hu Han (Han Hu) has accumulated a lot of achievements in academic and industrial experience —

According to Hu Han's personal academic homepage, he graduated from Tsinghua University with a bachelor's degree in 2008.

He obtained his doctorate in 2014, and then worked as a researcher at Baidu Institute of Deep Learning and Microsoft Research Asia successively.

In December 2023, Hu Han joined Tencent Hunyuan as a distinguished scientist, leading the team of vision-language model (VLM) and embodied intelligence foundation models.

During his tenure at Tencent, he also served as the technical lead of the Hunyuan large language model, and established the Hunyuan speech team.

In terms of academic research, one of Hu Han's most representative achievements is Swin Transformer in the field of computer vision.

In 2021, as one of the co-contributors of the paper, Hu Han and his team proposed a hierarchical vision Transformer architecture based on shifted windows.

This research improves the computational efficiency of Transformer when processing high-resolution images, enabling it to be better applied to various visual tasks such as image classification, object detection, and semantic segmentation.

With this achievement, the team won the Best Paper Award at ICCV 2021, which is also one of the representative academic honors in the field of computer vision:

In addition, Hu Han was selected into the 2022 MIT Technology Review China "35 Innovators Under 35" list.

Looking back at his research experience, we can see a relatively clear technical path —

In the early stage, he mainly focused on computer vision and vision foundation models, then gradually moved to multimodal large models, and further participated in the R&D of embodied intelligence foundation models during his time at Tencent.

These research directions are strongly related to the visual perception, environmental understanding and task execution capabilities required by robotics AI, which also provides technical accumulation for him to found Winverse AI.

As for financing, according to Bloomberg, DeepSeek plans to lead the multi-million-dollar seed round financing of Winverse AI, and Longzhu Capital under Meituan has also made an investment commitment.

If the transaction is successfully completed, it will be a relatively large investment in a startup by DeepSeek in the more than three years since its establishment.

There is another noteworthy background for this financing —

It is understood that Tencent is also one of the potential investors in DeepSeek's new round of financing. Now, the robotics AI company founded by the former Tencent scientist has received attention from DeepSeek and Meituan-affiliated capital.

From large model technology R&D to embodied intelligence entrepreneurship, the talent and capital links between leading technology companies have begun to extend more to the robotics field.

For DeepSeek, this investment also continues its layout in the robotics sector since the beginning of this year.

From Financing to Investment, DeepSeek Steps Up Layout in Robotics

Since the beginning of this year, DeepSeek's moves in the capital market have increased significantly.

In June this year, DeepSeek completed its first round of external financing since its establishment.

According to reports, the size of its first round of external financing is about 7.4 billion US dollars, ending the stage of long-term reliance on HFQuant for financial support.

Two months later, DeepSeek appeared on the IPO strategic placement list of Unitree Robotics.

According to the strategic placement results announced by Unitree Robotics on August 6, DeepSeek was allocated about 933,400 shares. Calculated at the issue price of 150.80 yuan per share, the subscription amount is about 141 million yuan, and the share lock-up period is 36 months.

In this placement, DeepSeek participated as an industrial strategic investor, and also signed a strategic cooperation memorandum with Unitree Robotics.

The two sides plan to carry out joint R&D around AI large models and embodied intelligence, focusing on exploring how to improve robots' ability to understand complex environments and generalize when facing different tasks and scenarios.

At the specific business level —

When DeepSeek carries out exploration related to embodied intelligence, it will give priority to choosing Unitree's robot products and technical solutions under the same conditions. When Unitree carries out model adaptation, training and integrated development, it will also give priority to DeepSeek's technical services.

In addition, DeepSeek will provide support in model architecture, intelligent computing cluster construction, data center operation and other aspects according to actual needs.

In October, new progress was made in DeepSeek's own financing and listing plans.

According to Bloomberg's report on October 6, DeepSeek is raising at least 800 billion yuan in a new round of financing, the final size may be close to 1 trillion yuan, with a target valuation of about 5 trillion yuan.

Tencent, CATL and other enterprises also participated in it. An important background for this financing is that DeepSeek is preparing to launch its IPO in early 2027, and the listing location may choose the STAR Market, which has not been officially determined yet.

Now, two months later, DeepSeek has set its sights on Winverse AI, which specializes in robotics foundation models.

The technical directions covered by the two investments are also different.

The technical accumulation of Unitree mainly focuses on robot ontology, motion control and hardware platforms, which can provide an operating carrier in the real world for embodied intelligence technology.

Winverse AI focuses its R&D direction on robotics foundation models, which is closer to the environmental perception, understanding and decision-making capabilities of robots.

One focuses on the "body" of the robot, while the other targets the "brain" of the robot.

At present, DeepSeek has not disclosed whether there is a technical cooperation plan between the two enterprises, nor has it announced independent R&D of robot products.

However, with successive investments in robot ontology enterprises and embodied intelligence model teams, DeepSeek's layout in the robotics sector has begun to take shape.

Reference Links:

[1]https://ancientmooner.github.io/

[2]https://www.bloomberg.com/news/articles/2026-10-12/deepseek-backs-ex-tencent-ai-scientist-s-400-million-robot-lab

This article is from the WeChat official account "QbitAI", author: Meng Yao, published with authorization from 36Kr.