Tian Keyu's new company does not have a name yet, but it is already valued at $200 million. Bloomberg reported on October 7, 2026 that the former ByteDance intern's world-model laboratory has raised close to $30 million from backers including 5Y Capital and IDG Capital, at a post-money valuation of about $200 million. The venture has no public product and has not announced a name.
What investors are actually buying
World models aim to teach systems how physical environments appear and change, with potential uses pointing toward robotics, interactive video and autonomous driving. The report says the team is building a visual language system of roughly 200,000 symbols, plans to train on about 100 million hours of video, targets cutting the cost of generating one second of video to a tenth or less of current levels, and intends to demonstrate the technology before releasing a full model in 2027. All of that remains a plan: there is no verifiable product or benchmark result yet, so the funding essentially buys time for a team and a direction.
Why this founder's record commands that price
Tian is the first author of Visual Autoregressive Modeling, an image-generation paper that won recognition at NeurIPS. His better-known episode came during his ByteDance internship, when the company said he had tampered with code and interfered with model training; ByteDance dismissed him and sued for 8 million yuan in damages, a case Reuters reported in 2024. That controversy has not stopped fresh capital from betting on him — a sign of how hot the world-model direction has become, with star projects such as Fei-Fei Li's World Labs pushing valuation anchors for the whole field higher.
Too early to call it a race against the star labs
There is still no agreed benchmark for world models, and a $200 million valuation here prices a research-stage team, not revenue or a product. The field also includes big technology companies and celebrity laboratories spending heavily, and a nameless lab will not find it easy to keep up on compute and data. The next observable milestones are whether the promised technology demonstration arrives on schedule and whether the 2027 timeline for a full model holds. Until then, this round looks more like an expensive ticket into the field than a verdict on who wins it.
