Tsinghua-MIT Startup TensorWorld Says It Raised a $105 Million Seed Round to Build Physics World Models

A three-month-old Beijing startup has claimed one of the largest seed rounds the physical world model track has seen. Wuyuan Technology, which operates under the brand TensorWorld, announced on X and through Chinese media coverage that it has closed a $105 million seed round, roughly 705 million yuan, to build what it calls physics world models: systems that predict how the physical world responds to actions rather than merely how it looks.

The investor list mixes financial and strategic names: Cathay Capital, HongShan (Sequoia China), Ant Group, MSA Capital, Meituan, IDG Capital, CDH VGC, Xinglian Capital, Fortune Capital and NIO Capital. The presence of Ant Group, Meituan and NIO Capital is telling — local-life services, mobility and vehicle businesses all have reasons to want models that can anticipate physical interactions, from a delivery robot navigating a kitchen to a car anticipating how cargo shifts.

Founder and CEO Wu Haixu was born in 1998 and completed his PhD at Tsinghua University's School of Software under Long Mingsheng, winning Tsinghua's special scholarship in 2022, before a postdoc at MIT CSAIL under Wojciech Matusik. His academic work centers on physical system modeling: Transolver, a physics-system solver, and GeoPT, which uses synthetic dynamics to lift static geometry into dynamic space, letting models learn particle trajectory evolution without labeled data.

The company draws a clear line against the dominant video-generation approach. Where video models chase photorealistic, temporally coherent footage, TensorWorld says it targets physical predictability: given an action, the model should judge how an object will move, collide and deform, and transfer that judgment to unfamiliar scenes. That maps directly onto what embodied-AI teams describe as their core bottleneck — robots operating in real environments lack physical intuition, and sim-to-real transfer remains expensive and brittle.

The founding team covers three dimensions the company says are complementary: three-dimensional understanding, reinforcement learning and agent planning, plus large-scale compute. Co-founder Zhou Junsheng contributed to Uni3D, Weirui Ye proposed EfficientZero, and Wen Zheng brings experience with compute infrastructure exceeding 100,000 accelerators.

The caveats are as important as the headline. There is no agreed public benchmark for physical world models, and the company's figures — round size, investor participation, team credentials — come from its own announcement and Chinese media coverage, not from independent verification. The GeoPT approach is academically sound, but the engineering distance between a paper's synthetic dynamics and a trainable, scalable model is unknown, and data requirements could be enormous.

The validation test is not the size of the check. It is whether TensorWorld's models can output measurable prediction accuracy in robotics or industrial simulation settings, and whether that accuracy keeps improving with scale. Until then, the valuation logic in this track rests on talent and narrative — and a $105 million seed is a very expensive narrative bet on physics being the missing piece of embodied AI.