Most model labs train one enormous network and hope it generalises. A Shenzhen startup backed by well-known Chinese technology investors is betting on a stable of smaller specialists that pass work to each other.
Hypersphere Technology has raised hundreds of millions of RMB in angel funding, per an October 26 report. Legend Capital, CAS Star, SenseTime Guoxiang Capital and WestSummit Capital participated.
Founder Weiyang Liu, a professor at the Chinese University of Hong Kong, calls the approach Infinite Models. Instead of a monolith, it binds its own expert models to capabilities drawn from open-weight and proprietary systems, using routing and aggregation to match the right model to each request.
The engineering bet behind routing
Routing is no longer exotic. Sending simple queries to small models and hard ones to large models can slash inference bills, a point enterprises now press on as token costs get audited. The difficulty sits in reliability: a router has to keep quality steady when a third-party model behind it is deprecated, repriced or withdrawn.
Hypersphere’s research spans pretraining, reinforcement-learning post-training, parameter-efficient fine-tuning and model merging, the tooling needed to keep a family of specialists coherent.
China’s market is crowded. MiniMax, Moonshot and DeepSeek all ship competitive systems, several with open weights, leaving a newcomer to justify fresh capital in a market already arguing over price wars.
The angel round is a vote that routing and merging can become a defensible product rather than a tuning trick. Hypersphere has not named enterprise customers.
