Tian Keyu raises nearly $30M for a world-model lab valued at $200M
The former ByteDance intern's research venture has backing from 5Y Capital and IDG Capital, but Bloomberg reports it has no public name or product.
By RuntimeWire Staff · Published
Primary source: Bloomberg Technology
Why it matters
Tian's round values a research-stage world-model lab with no public product at $200 million after the investment. The lab's next test is whether his image-generation research record can translate into models that represent and predict physical environments.
Tian Keyu has raised close to $30 million for a new AI research lab pursuing world models, with investors valuing the unnamed venture at $200 million after the investment, Bloomberg reported on October 7th. The backers include 5Y Capital and IDG Capital. Bloomberg says the lab has no public name or product.
The financing gives Tian a chance to build in a field where researchers aim to make AI systems represent how physical environments look and change. His research record in image generation provides a starting point. The lab would need to model space, motion and physical behavior well enough to support applications such as robotics, interactive video or autonomous driving. Bloomberg reported those as potential uses, not products the lab has demonstrated.
From image generation to worlds
Tian completed a master's degree at Peking University in summer 2025, according to Bloomberg. Before starting the lab, he was the first author of Visual Autoregressive Modeling: Scalable Image Generation via Next-Scale Prediction, a 2024 paper from researchers affiliated with Peking University and ByteDance. NeurIPS named the paper a Best Paper. Its method generates images from coarse to fine, predicting at successively higher resolutions rather than producing a sequence of image patches. The paper and award establish Tian's research credentials, but do not show that the new lab has solved world modeling.
A convincing still image is different from a coherent environment that changes as a camera moves, objects interact or an agent takes an action. A model meant to help train a robot or simulate a drive must represent how a scene looks and anticipate how it changes. Bloomberg has not reported a Tian lab model, demo, customer or benchmark, so the venture's technical progress cannot yet be compared with that standard.
With no public name or product, the $200 million post-investment valuation is an early bet on Tian and his ability to assemble a lab. Bloomberg did not identify a lead investor or break down the investors' stakes. It also did not report revenue, customers or a financing close date. The round's structure and the business milestones investors expect remain undisclosed.
A crowded field with public proof points
Bloomberg frames Tian's ambitions against Fei-Fei Li. Li's World Labs has released Marble, a model for generating navigable 3D worlds, and a public API for creating those environments from text, images and video. In February, World Labs said it raised $1 billion in new funding.
World Labs also announced on September 28th that it had signed a definitive agreement to join AMD, subject to regulatory approvals and other closing conditions. The deal puts Li's effort in a different institutional setting. Tian's plans and competitive approach remain undisclosed. RuntimeWire reported in August that Li's World Labs was pursuing persistent 3D environments for creators, designers and robotics researchers. The lab's announced AMD agreement could make access to hardware and research infrastructure part of the competition alongside model quality.
World Labs' public funding announcement and product releases contrast with Tian's pre-product financing. The disclosed $30 million figure says little about compute capacity, staffing or how long the lab can train models. It also does not show whether the team will prioritize 3D generation, simulation, robotics or another application. Bloomberg describes a broad research ambition, not a go-to-market plan.
Tian's prior work gives investors a basis for the technical bet: he helped develop a recognized visual-generation method while affiliated with both academia and ByteDance. The new lab's work will provide the next evidence of its progress. A world model that can produce plausible scenes is one thing; a model that predicts physical changes reliably enough for downstream systems makes a much larger claim. Until Tian's lab publishes a system or demonstrates what it can do, the $200 million valuation reflects the opportunity and the team, not a verified capability.