Two Beijing artificial intelligence start-ups have unveiled rival five-level roadmaps for “world models” a day apart, as both attract heavy funding and speculation over possible Hong Kong listings. The scales look similar. They do not measure the same thing.
Zhu Jun, the Tsinghua University professor behind ShengShu Technology, presented his framework at the World Robot Conference on August 19 and published a technical report the following day. GigaAI founder Huang Guan set out his five levels at an August 20 forum, according to a company WeChat post dated August 27.
A world model tries to represent how an environment changes and how actions affect it. Where a language model predicts the next word, a world model is meant to predict the next state: how a cup tips, a robot arm collides or traffic changes after a car turns. The label now stretches across video generation, robot control and simulated environments, which is part of why rival definitions have appeared. Its proponents say the technology could allow robots and autonomous vehicles to learn in simulation before training in the slower, costlier real world. ShengShu describes it as a feedback loop joining understanding, prediction and action, rather than simply a better video generator.
GigaAI’s scale focuses largely on the world itself. Its five levels move from visually convincing video and executable actions to consistent geometry, accurate physics and, finally, environments that can maintain state and memory over long periods. The framework asks the model to satisfy progressively tougher constraints.
ShengShu’s ladder instead tracks the role of the machine inside that world. It begins with generation, moves through real-time interaction and robot action, then ends with autonomous agents and systems that co-ordinate robots, software and people. ShengShu maps its Vidu video generator, interactive Vidu S1 model and Motubrain robotics system onto Levels 1 to 3. It says long-term memory, open-ended exploration and continuous learning remain unresolved.
The mismatch is clearest at Level 2. For ShengShu, it is an interactive world that changes when a user speaks or acts. For GigaAI, it is the generation of executable actions intended to complete a task. The shared numbering makes the frameworks easy for investors to compare — and easy to confuse: a higher number on one scale does not put a company ahead on the other.
Prominent overseas developers have so far framed progress differently. Nvidia released Cosmos 3 as downloadable open models in May. Google began offering Project Genie, an experimental world-building prototype, to Google AI Ultra subscribers in the US in January. World Labs’ June taxonomy divides world models into renderers, simulators and planners rather than levels of maturity.
The Chinese frameworks have arrived during a rush of capital into the sector. Local media have cast each as a contender to become China’s first listed “world-model” company. Reports say GigaAI has raised more than Rmb3.5bn ($490mn) since March, while Huang has said an ongoing round could value it at as much as $3bn. The company said in July that reports of a Hong Kong listing as soon as 2026 misread his remarks and were not official information.
ShengShu raised nearly Rmb2bn in April and a further Rmb3.4bn in July, according to Chinese media. Bloomberg reported on August 18 that the company was considering a Hong Kong offering that might not take place until 2027.
Neither ladder has an outside referee. Both companies define their own rungs, and ShengShu grades itself against them. The comparison with autonomous driving is instructive: the widely used SAE scale was written by an engineering standards body, not a carmaker. For investors, the first question is not which company has climbed higher, but which ladder they are looking at.
