Realtime AI News
Liang Wenfeng Backs Swin Transformer Author as DeepSeek Leads Winverse AI Round
DeepSeek plans to lead a seed round of tens of millions of dollars in Winverse AI, a robotics-AI startup founded in 2026 by Hu Han, a co-author of the influential Swin Transformer and a former Tencent Hunyuan distinguished scientist. The deal would value the company at about $400 million and marks DeepSeek's second robotics investment in two months, after taking a stake in Unitree Robotics.
Liang Wenfeng is placing another bet on robotics. According to a QbitAI report citing Bloomberg, DeepSeek plans to lead a seed round of tens of millions of dollars in Winverse AI, a robotics-AI startup, at a post-money valuation of about $400 million, or roughly 2.8 billion yuan. Meituan's Longzhu Capital has also committed to invest. If the deal closes, it would rank among the largest startup investments in DeepSeek's three-plus years of existence.
Winverse AI was founded in 2026 and is still in early R&D, with no public details yet on its products or technical roadmap. The draw is its founder. According to his academic homepage, Hu Han (Han Hu) earned his bachelor's degree at Tsinghua University in 2008 and his PhD in 2014, then researched at Baidu's Institute of Deep Learning and Microsoft Research Asia.
In December 2023, Hu joined Tencent Hunyuan as a distinguished scientist, leading teams working on vision-language models and embodied-intelligence foundation models; he also served as technical lead for Hunyuan's language model and built its speech team. His best-known academic contribution is Swin Transformer, a hierarchical vision Transformer architecture based on shifted windows that he co-authored in 2021.
Swin Transformer improved the computational efficiency of Transformers on high-resolution images, making them more usable for image classification, object detection and semantic segmentation. The work won a Best Paper award at ICCV 2021, and Hu was named to MIT Technology Review's 2022 "35 Innovators Under 35" list for China. His path, from computer vision to multimodal models to embodied foundation models, maps closely onto the perception, understanding and task-execution capabilities that robotics AI needs.
It is DeepSeek's second robotics investment in two months. In August, DeepSeek appeared as an industrial strategic investor in Unitree Robotics' IPO allocation: Unitree disclosed on Aug. 6 that DeepSeek was allotted about 933,400 shares at an issue price of 150.80 yuan, an investment of roughly 141 million yuan with a 36-month lock-up. The two also signed a strategic cooperation memorandum to jointly develop AI large models and embodied intelligence.
The two investments target different layers. Unitree's strength is in robot bodies, motion control and hardware platforms, the robot's "body," while Winverse AI focuses on robot foundation models, closer to perception, understanding and decision-making, the robot's "brain." DeepSeek has not disclosed whether the two companies will cooperate technically, nor has it announced any robot product of its own.
DeepSeek has become far more active in capital markets this year. It closed its first external funding round in June, reportedly around $7.4 billion, ending its long reliance on High-Flyer Quant for capital. Bloomberg reported on Oct. 6 that DeepSeek is raising at least 80 billion yuan in a new round that could reach nearly 100 billion yuan at a target valuation of about 500 billion yuan, with Tencent and CATL among the participants, as the company prepares a possible IPO as early as 2027.
Tencent is itself a potential investor in DeepSeek's new round, making it notable that a robotics-AI company founded by a former Tencent scientist is drawing interest from DeepSeek and Meituan-linked capital. The pattern, of talent and capital ties among leading tech companies widening into robotics and embodied AI, is the signal worth watching. Next: whether the seed round closes, whether the two sides announce technical cooperation, and whether DeepSeek moves toward building its own robots.
Why it matters
The move shows DeepSeek extending from model research into the embodied-AI stack, pairing a "brain" investment in Winverse AI with its earlier stake in robot-maker Unitree. If more Chinese model labs and robotics firms pair up this way, the race in embodied intelligence could consolidate quickly around a few model-plus-hardware alliances.
Nearby Updates
All10/13, 18:00
Open models have real benefits. Do they outweigh the risks?
Open models have real benefits. Do they outweigh the risks?. MIT Technology Review Explains: Let our writers untangle the complex, messy world of science and technology to help you understand what’s coming next. You can read more from the series here. Late last month, Anthropic released a report arguing that the model G...
10/13, 18:00
Google Pulls Its Best Gemini Models From Free Users; Android Police Points to a Cheaper Fix
Android Police reports that Google has removed its most capable Gemini models from the free tier, so free accounts no longer get the flagship Gemini experience. The outlet also lays out the lowest-cost workaround for users who still want access to the top models.
10/13, 17:00
AI labs are buying data from dead startups to train their models
AI labs are buying data from defunct startups to train their models, MIT Technology Review reports. One founder who wound down his crypto startup turned its digital archives into roughly $25,000 for the code base alone, plus a still-pending sum.
10/13, 16:48
Kakao Plans to Turn KakaoTalk Into an AI Agent Platform
Kakao plans to turn KakaoTalk, its flagship messaging app, into a platform for AI agents, according to a report by THE ELEC. If it lands, one of Korea's most-used chat surfaces would become a major distribution channel for agents that live inside everyday conversations.