Chinese investors are paying up early for academic founders: 49 of the 182 AI companies founded and funded in 2026 by early October came from universities or research institutes.
Hypersphere, a Shenzhen startup building large AI models, has closed an angel round worth hundreds of millions of yuan, according to a People’s Financial News brief carried by Eastmoney on October 26. The investors are Legend Capital, CAS Star, SenseTime Guoxiang Capital, WestSummit Capital and Gengxin Capital China. Jiwei, the chip-industry outlet that first reported the round, added an unnamed top-tier US dollar fund and said the company was founded by Weiyang Liu, a professor at the Chinese University of Hong Kong (CUHK), as did TechNode. No exact amount or post-round valuation has been disclosed.
The company’s bet is on many models working together. It calls its approach "Infinite Models," pitched by Jiwei as the next way to scale AI after ever-larger single models. Hypersphere plans to train its own expert models and also draw on outside models such as DeepSeek, Kimi, GPT and Claude, using routing, which picks the model that handles each request, and aggregation, which merges several models’ answers.
The round is large for a company only weeks old. Hypersphere was registered on August 5, 2026, and had fewer than 20 staff, all holding doctorates from leading overseas universities and research institutes. Those details come from an October 10 profile by ChinaVenture, a Chinese venture-capital news outlet. The company has not said how it will spend the money or how much it raised.
A Hong Kong academic team draws a broad group of investors
ChinaVenture described Liu as an assistant professor and doctoral supervisor at CUHK, where he leads a research group called SphereLab; Jiwei and TechNode call him a professor. Since 2025 he has also been an assistant professor at Shenzhen’s Hetao College, the profile said. He earned a computer science doctorate at the Georgia Institute of Technology and a machine learning doctorate at the University of Cambridge, then did postdoctoral work at Germany’s Max Planck Institute under the machine learning researcher Bernhard Schölkopf. He has also worked at Google Brain, now part of Google DeepMind, and at NVIDIA, according to Jiwei.
Jiwei counts more than 80 papers at leading AI venues and over 14,000 Google Scholar citations, and says MetaMath, a math dataset his team built, was used in Meta’s Llama 3. ChinaVenture noted a 2023 best paper award from the IEEE Signal Processing Society. Around 2017 he proposed face recognition methods built on spherical geometry, which the profile suggested may explain the company’s name. Since 2023 his work has moved to diffusion and language models, using orthogonal transformations, a class of mathematical operations, to constrain how a model’s parameters change during training.
The investor list mixes institutional venture capital with an AI company’s own fund. Guoxiang Capital belongs to SenseTime, ChinaVenture said, and the People’s Financial News brief gives WestSummit Capital’s Chinese name as 华山资本. TechNode’s Chinese edition also lists an unnamed US dollar fund and said Gengxin Capital China acted as the company’s exclusive and long-term strategic financial advisor.
Hypersphere’s research spans pretraining, fine-tuning and model merging
TechNode’s English report listed Hypersphere’s research areas as pretraining, reinforcement-learning post-training, parameter-efficient fine-tuning and model merging. Jiwei named the tools behind them: Orbit, a post-training framework for trillion-parameter models that it says has been validated on Kimi-K2.6 and DeepSeek V4; POET-X, a pretraining framework; and OFT, a fine-tuning method that adjusts only a small share of a model’s parameters.
The Chinese edition of TechNode said the company would train specialized models on top of foundation models and have them cooperate, while continuing to scale up the training of individual models. That second point matters for investors, because it means Hypersphere still plans to build foundation models of its own rather than act only as a layer that coordinates other companies’ systems.
The round also fits a wider pattern. ChinaVenture counted 182 companies in China’s AI industry chain founded since January 1, 2026, that had raised money by the time of its analysis. Of those, 33 were started by university or research-institute faculty and researchers, and 16 came from technology transfer or teacher-student teams, a combined 49, or about 27%. ChinaVenture had already profiled Hypersphere as one of its examples, noting that the funding was first disclosed at the end of September. At that stage the company claimed backing from a top-tier US dollar fund and a valuation above 100 million. The angel round comes after China’s AI funding cooled in August, when robot makers led the deal count.
A training framework for expert models comes next
According to TechNode’s Chinese edition, Hypersphere plans to keep developing a training framework for specialized models and to work with industry partners on training under different data sets and tasks. Its long-term goal is a foundation model designed for building specialized models.
None of the reports names a product, a release date or a customer. Liu had been scheduled to give a technical talk at the GOSIM conference in Shenzhen on efficient training of large language models with orthogonal transformations, ChinaVenture said, which offers an early look at the methods the company may use. The first public test will be a released expert model or a working routing system that can be measured against single large models on cost and accuracy.