DeepSeek Weighs a Move Into Robotics as Liang Wenfeng Meets Founders
DeepSeek founder Liang Wenfeng has been holding talks with founders working on embodied intelligence and is weighing whether the company should move into robotics, according to a report from Phoenix Tech. People familiar with the matter said two routes are under discussion, with one being a dedicated internal team and the other a set of stakes in companies already working in the field. DeepSeek has not commented.
The conversations covered robotics, large models and the way AI systems interact with the physical world. One person told Phoenix Tech that Liang does not have deep experience in the area and has been studying it intensively. No decision has been reported, and the company had not responded to requests for comment at the time of publication.
But this is not DeepSeek's first step towards machines. Earlier this year the company took a strategic placement in Unitree Robotics worth about 141 million yuan, or roughly 21 million US dollars, for a stake of about 2%. The two signed a memorandum covering joint work on large models and embodied intelligence, under which Unitree gives priority to DeepSeek's technical services and DeepSeek gives priority to Unitree's robots.
Its more recent move points at the software side of the problem, with DeepSeek planning to lead a seed round of tens of millions of dollars in Winverse AI, a company founded this year by Hu Han, a former Tencent distinguished scientist who led vision and language and embodied intelligence model teams there. The round would value Winverse at about 400 million dollars, and Longzhu Capital, a vehicle tied to Meituan, has also made a commitment. The deal is not final and the terms could still change.
Some in the industry question how close that investment sits to robotics proper. Winverse AI works on model acceleration and other underlying infrastructure, which makes it a supplier to embodied intelligence companies rather than one of them. The distinction matters when the money is being described as a robotics bet.
But part of the thinking appears to be competitive. People close to DeepSeek said OpenAI's GPT-6 Astra reached a 95% success rate on a robot control task, against about 40% for the model it followed, and that the demonstrations surprised the industry, while Liang has said publicly that DeepSeek still trails GPT-6 on some measures. One industry figure cautioned that even the stronger model needs about 40 minutes to pick up a cup, so the bar for useful robots remains low.
DeepSeek's own work has stayed on the software side. It has concentrated on language models and the tooling around them, and its recent releases include an open source agent runtime and versions of its inference libraries built for domestic chips. Liang has said the company's next models must be trained on Chinese silicon.
The company is also preparing to sell shares. According to Bloomberg, it completed a first external funding round in the middle of the year and is working towards a stock market listing. Tencent and CATL have been named among the investors in its latest round.
So rivals are moving too. Zhipu and Moonshot AI, the company behind Kimi, both hold embodied intelligence work in reserve, while Skild AI and Physical Intelligence in the United States are building general robot brains. China leads on robot hardware and the Unitree listing this year underlined that position, so what none of this settles is whether DeepSeek will build robots, back the companies that do, or stay where it is.
DeepSeek founder Liang Wenfeng has been meeting embodied intelligence founders and is weighing a move into robotics, according to Phoenix Tech, with no decision reported
DeepSeek already holds about 2% of Unitree Robotics after a placement worth about 141 million yuan earlier this year
It plans to lead a seed round in Winverse AI at a valuation of about 400 million dollars, in a deal that is not final
One reported driver is OpenAI's GPT-6 Astra, which people close to DeepSeek said reached a 95% success rate on a robot control task
Source: Pandaily
