Dexmal (原力灵机), a Chinese embodied intelligence startup founded in March 2025, has completed a new st...

Dexmal (原力灵机), a Chinese embodied intelligence startup founded in March 2025, has completed a new strategic funding round with participation from major domestic AI labs including Zhipu AI (智谱), Stepfun (阶跃星辰), SenseTime (商汤科技), and Alibaba (阿里巴巴). The company, whose core team originates from Megvii (旷视科技), claims to have pioneered an "embodied-native" approach that rebuilds AI systems from scratch for physical environments rather than adapting traditional AI models. Its flagship achievement, DM0, is billed as the world's first embodied-native large model, combining robot perception data, autonomous driving data, and web data for joint training, achieving sub-millimeter precision and winning first place in both single-task and multi-task categories at the international RoboChallenge benchmark. Concurrently, Dexmal announced it has merged via equity acquisition with Atomix, a logistics robotics firm also spun out of Megvii's internal smart logistics division, which ranks second globally in pallet shuttle vehicle sales with over 500 deployed projects.

Why this matters: Dexmal's backing by four of China's most prominent foundation-model labs signals a strategic bet on embodied AI as the next frontier for large-model distribution, moving beyond language and vision into physical action. The merger with Atomix exemplifies the acqui-licensing pattern — combining algorithmic innovation with proven hardware deployment and real-world logistics channels, creating a vertically integrated stack from model to robot to commercial application. This mirrors recurring pattern 5.3 (hyperscaler-distribution moat) but with a physical-world twist: instead of cloud distribution, the moat is built through installed robotic hardware and logistics contracts. The involvement of Zhipu, Stepfun, and SenseTime — all competing foundation-model players — suggests coordinated joint investment in a shared physical-AI platform, rather than exclusive ownership.

Expert take: Dexmal's approach represents a deliberate departure from the prevailing "AI-on-robot" paradigm, where general-purpose models are bolted onto existing hardware. By claiming the "embodied-native" frame and backing it with a merged entity that has real logistics revenue, Dexmal is attempting to compress the capital cycle for embodied AI: bypass the years of hardware deployment that rivals like Figure or 1X have endured. The July product roadmap — DM0.5, a general-purpose robot, and new application infrastructure — will test whether this accelerated stack can deliver general-purpose manipulation in commercial settings. The key open question is whether the logistics vertical provides enough task diversity to train truly general-purpose physical models, or whether it will produce a narrow specialist.

#EmbodiedAI #ChinaAI #Robotics #FoundationModels #AcquiLicensing #LogisticsRobotics