MarsBit, citing a report from the WeChat account IT Juzi, said at least nine AI startups in China and overseas had crossed the unicorn threshold within six months of being founded as of early September 2026. The list spans China’s Yuyong Technology, Kunlunxing Robotics, AGILINK and Naive.ai, as well as River AI, Hark, AMI Labs, Recursive Superintelligence and Atoms abroad. The report argues that many of these companies were funded before products, revenue or commercial validation were in place, with capital instead assigning value to founder track records, technical direction and ecosystem positioning. Cases highlighted include Yuyong Technology reaching an about $2 billion valuation roughly three months after registration, AGILINK becoming the only company in the group already generating revenue, AMI Labs raising what the report described as Europe’s largest seed round, and River AI securing $1.1 billion within about two months of surfacing. The piece says the current market is rewarding scarcity first and waiting for proof later, while warning that delivery risk, ecosystem dependence and lofty expectations could make the next 12 to 24 months decisive for these companies.

Huawei alumni have become one of the strongest fundraising groups in China’s embodied intelligence sector this year, according to incomplete data cited from ITjuzi. From Jan. 1 to Aug. 14, 2026, 12 embodied AI companies founded or co-founded by entrepreneurs with Huawei backgrounds completed 24 financing rounds, with total funding estimated at roughly 8.635 billion yuan. The pace was heavily concentrated in early stages, with 9 angel rounds, 5 seed rounds and 5 Pre-A rounds, meaning more than 80% of the deals landed before companies had spent much time in market. Capital was also concentrated at the top. X Square Intelligence, founded less than two years ago by former Huawei autonomous driving CTO and chief scientist Chen Yilun and former Baidu intelligent driving executive Li Zhenyu, closed a $455 million Pre-A round in April, described in the report as the largest single financing round in China’s embodied intelligence industry. Kunlunxing Robotics also raised funding worth tens of billions of yuan, and the two companies together accounted for about 70% of the Huawei-linked total. Other names highlighted in the report include Ola Wanxiang, Knowin Intelligence, Gunao Panshi, Beta Infinite, Moxi Intelligence, Moushen Intelligence, Yuansheng Intelligence, Xinzhi Jushen, Zhicheng AI and Qiwu Technology.

RoboParty, an open-source bipedal humanoid robotics startup founded by 22-year-old Huang Yi, has completed back-to-back angel++ and Pre-A financing rounds totaling nearly 500 million yuan, with battery giant Contemporary Amperex Technology Co. Limited (CATL) serving as the sole investor in the Pre-A round. The company was formed in Shanghai in 2025 after Huang graduated early from Harbin Institute of Technology and moved with classmates to Zhangjiang Robot Valley. According to the report republished by MarsBit from Chinese media outlet PEdaily, RoboParty positions itself as a full-stack embodied intelligence platform, combining open-source humanoid bodies, in-house hardware development, mass-production preparation, foundation-model work and an operating layer called Party OS. Its RPO open-source humanoid project, released in January, has drawn nearly 10,000 followers in the global developer community, more than 2,000 GitHub stars and close to 1,000 orders from developers, universities, research institutions and model teams. The company said RP1 is scheduled to launch in the fourth quarter. The report also said RoboParty has completed six financing rounds in eight months and seen its valuation rise by more than 20 times, while investors now include Matrix Partners China, Xiaomi Strategic Investment, Galaxy General, SenseTime Guoxiang Capital, Baidu Ventures and others. CATL’s direct investment through its group strategic investment arm was described as an unusual move and a potential signal of industrial-chain collaboration.

VISReg, a new self-supervised learning method built around Variance-Invariance-Sketching Regularization, has drawn repeated reposts from Turing Award winner Yann LeCun, who described the line of work as: "VICReg begat SIGReg which begat VISReg." The paper positions itself as a response to one of the central problems in JEPA-style world models: representation collapse, where distinct inputs are mapped to the same or only a few vectors and the model stops learning useful features. The method splits anti-collapse regularization into two independent targets, scale and shape. It keeps a variance term to control scale, replaces covariance constraints with a sketching objective based on Sliced Wasserstein Distance to capture distribution shape, and uses stop-gradient to separate the two during optimization. According to the paper, this avoids the vanishing-gradient issue seen in SIGReg when collapse begins. The authors report results across 15 datasets, including 8 in-domain benchmarks, 6 out-of-distribution datasets, and ADE20K dense prediction. The paper says VISReg outperformed seven mainstream self-supervised methods in aggregate and matched DINOv2 on OOD benchmarks using about one-tenth of the training data. Code, pretrained weights, and the paper are publicly available.

HWM introduces hierarchical planning to world models, addressing error accumulation and rising search costs in long-horizon tasks. In real-world tests, it achieved a 70% success rate versus 0% for single-layer models.

HWM introduces hierarchical planning to improve long-horizon task execution in world models. In real-world object handling tests, it achieved a 70% success rate versus 0% for single-layer models, while also reducing planning costs in some simulated environments.