The World Model
Category: Foundation Models / LLMs
A stealth Beijing AI lab founded by VAR author Keyu Tian. It is building an autoregressive world foundation model that learns real-world physics from video. The lab has no official name yet, so 'The World Model' is a descriptive label. The company is led by Keyu Tian. Based in Beijing, China. Team size: 1-10. Total funding raised: $30M. Latest round: Undisclosed. Key investors include 5Y Capital (Wuyuan Capital), IDG Capital.
- Headquarters
- Beijing, China
- Team size
- 1-10
- Total funding
- $30M
Value proposition
Trains AI to understand physical rules, predict object motion and model how environments change with actions. The lab claims its approach can cut the cost of generating one second of video by at least an order of magnitude.
Products and solutions
No public product yet. The lab is building a world foundation model using a custom 'visual dictionary' of about 200,000 machine-only symbols (tokens) for video compression, representation and future-frame prediction. It plans to train on about 100 million hours of video, with a full model release planned for 2027.
Unique value
An efficiency-first, autoregressive, token-based world model that extends the VAR next-scale-prediction lineage. It competes on compute cost rather than on 3D spatial reconstruction.
Target customer
Future robotics, autonomous driving and interactive video developers. There is no commercial product yet, and the lab is focused on foundational research.
Industries served
Robotics, Autonomous Driving, Interactive Video / Generative Media
Technology advantage
The founder authored VAR (NeurIPS 2024 Best Paper). The lab has a proprietary 200K-symbol visual tokenizer that brings LLM-style scaling to video, and it targets the compute-cost bottleneck of world models.
How they differentiate
Uses autoregressive symbolic tokenization of video plus massive video pretraining, with a focus on compute efficiency. World Labs instead pursues 3D spatial intelligence and scene reconstruction.
Main competitors
World Labs (Fei-Fei Li; AMD acquisition announced September 2026, closing pending), Google DeepMind (Genie), other Chinese world-model startups (33+ in China per 2026 industry maps)
Key partnerships
Incubated by 5Y Capital (partner Meng Xing)
Major milestones
Jul 2026: Chinese media first reported the venture, incubated by 5Y Capital at about a $200M valuation. Oct 2026: Bloomberg reported a ~$30M raise from 5Y and IDG at $200M post-money. 2027: planned foundation model release.
Growth metrics
About a 10-person team; about $200M post-money valuation; no product or revenue yet.
Market positioning
A well-funded, pre-product research lab in the crowded 2026 world-model race. It is positioned on the founder's research reputation and an efficiency thesis.
Geographic focus
China-based, competing globally in world foundation models
About Keyu Tian
Ex-ByteDance Commercialization Technology research intern (2021-2024; dismissed Aug 2024; ByteDance sued him for RMB 8M over interference with model training, and he has since acknowledged serious misconduct); Peking University PhD student (advisor Wang Liwei); Beihang University BS Software Engineering. First author of 'Visual Autoregressive Modeling' (VAR), NeurIPS 2024 Best Paper; also authored SparK (Designing BERT for ConvNets).