Moushen Brain haalt bijna 100 miljoen yuan op in Pre-A-ronde: Fudan-professor en voormalig Intel-hoofdwetenschapper creëren 'brein' voor robots

Chinese startup 'Motion Brain' (眸深智能) has raised nearly 100 million yuan (approximately $13.8 million) in a Pre-A funding round. Investors included industry investment platform Jinyue Investment, co-founded by a major Chinese asset management company, a Hong Kong consortium, and several public companies, as well as fund Chuanghehui Capital and previous investor Xuhui Capital. This is the second round in two months, following a Pre-A round of 300 million yuan in May 2026. Additionally, a Pre-A+ round of 500 million yuan is in the process of closing. Since the beginning of the year, the company's valuation has increased more than tenfold. Motion Brain was founded in January 2025 by Professor Chen Tao, director of the Deep Learning Laboratory at Fudan University; former Intel Chief Scientist in China, Zhang Yimin; and serial entrepreneur Mu Zelin. Key employees have experience at Huawei HiSilicon, Intel, and Nvidia. The company has been building its technological foundation since 2022. It developed MLD (Latent Space Diffusion Model), which for the first time in the world mapped motions into latent space and applied a diffusion model to generate natural movements. In September 2023, MotionGPT was introduced—a model that breaks down human poses into approximately 3,000 'motion tokens' (similar to tokens in Large Language Models), enabling robots to perform new actions without specific training (Zero-Shot). By 2026, seven generations of models had been released. Based on these, STI-WM (Spatio-Temporal Integrated World Action Model) was created—a framework for long-term planning, closed-loop control, and physical interaction, using a ratio of 80% internet video, 10% motion capture data, and 10% real robot data, reducing the need for real data by 90% and improving accuracy to 99%. The T²MB (Task*Task Motion Brain) model, announced in March, allows robots to learn autonomously on-site without sending data to the cloud, increasing task execution accuracy by up to 25%. The company's work is already cited in Nvidia's latest models (ARDY). A technical advantage is the compression of models from hundreds of billions to tens of billions of parameters, reducing latency on the device from 200 to 10 milliseconds, lowering computation costs from 200,000 to 10,000 yuan per model, and adapting to Chinese chips (HiSilicon Ascend, Horizon Robotics, Enflame). The compression work earned the IJCAI 2025 Best Paper Award—the only one awarded to a Chinese mainland team in the past five years.