Yuanli Lingji raises Series A from Zhipu, Stepfun, SenseTime — acquires logistics robot maker Atomix
What's the deal? Yuanli Lingji, an embodied intelligence startup founded in March 2025 by Megvii co-founder Tang Wenbin, has closed a new funding round backed by three of China's leading AI model companies: Zhipu AI, Stepfun, and SenseTime. Industrial investors Huaqin and SAIC Hengxu also participated.
Alongside the raise, Yuanli Lingji acquired AtomixDealroom has a profile for this one. Try Dealroom →, a logistics robotics company that Tang had spun out of Megvii in July 2024. Atomix generates close to ¥1B in annual revenue, serves over 500 projects across more than 20 countries, and counts UniqloDealroom has a profile for this one. Try Dealroom →, Mixue, and CATL among its clients.
Including Alibaba, which exclusively led the A+ round, Yuanli Lingji has now assembled four of China's major AI model makers as backers — a first for the embodied AI sector.
Why now? The deal signals a broader shift in China's AI landscape: as competition among large language model makers moves from generating text to controlling physical actions, embodied AI has become the next frontier. Zhipu and Stepfun had barely invested in the space before this round.
The merger is designed to solve what insiders call a "data deadlock." Embodied AI models need real-world operational data to improve, but robots cannot enter real scenarios without capable models — a chicken-and-egg problem. Atomix's 500-plus live projects provide a continuous stream of picking and logistics data that can feed Yuanli Lingji's model training directly.
Yuanli Lingji has already released DM0, an embodied foundation model that blends internet data, autonomous driving physics, and robot manipulation data. With just 2.4B parameters, DM0 achieves sub-millimetre precision and can transfer across different robot form factors.
What could go wrong? The embodied AI sector still lacks a clear scaling law — the predictable relationship between compute, data, and model performance that drove LLM progress. Integrating a near-¥1B revenue hardware business with an early-stage model company also carries execution risk.
Competition is intensifying fast. ByteDance is aggressively recruiting embodied AI leaders from top startups, and US-based Skild AI just acquired Zebra Technologies' robotics automation unit.
The signal: Zhipu AI, Stepfun, and SenseTime are all corporate investors backing Yuanli Lingji — a rare alignment of rival AI model houses around a single embodied intelligence play. That China's leading LLM builders are collectively placing bets here, despite minimal prior exposure to the robotics sector, underscores a conviction that the value in AI is shifting from generating tokens to controlling physical actions — and that the startups which own both the model and the real-world data pipeline will set the pace.
Read more: 36kr.com
