Dexmal / Force Intelligence原力灵机

About

Dexmal was founded in March 2025 by Tang Wenbin, co-founder and CTO of Megvii. The core founding team (Fan Haoqiang, Zhou Erjin, Wang Tiancai) are Megvii veterans, mostly from Tsinghua's Yao Class, and the company has been called the 'most large-model-flavored' player in the embodied track. With 'Embodied Native' as its core philosophy, its products span the MMLA end-to-end multimodal embodied foundation model, the DM0/DM0.5 embodied-native foundation models, the DW0.5 embodied world model, the DFOL post-training framework, the DexOS robot operating system and the APEX robot body, plus the embodied industry's first MaaS service (billed by token). The open-source ecosystem includes the Dexbotic toolbox, the DOS-W1 open-source dexterous hand, and the RoboChallenge real-robot evaluation platform released jointly with Hugging Face. In June 2026 the company completed an equity-based merger with logistics robot company Atomix (nearly RMB 1B in annual revenue, 500+ projects across over 20 countries) — the first 'model × scenario' super-merger in China's embodied track — and closed a new strategic round, lining up four top LLM vendors: Alibaba, Zhipu, Stepfun and SenseTime. As of the April 2026 Series B, it had raised over RMB 1.2B within a year.

Technical Approach

Proposes the 'Embodied Native' philosophy: data, training and frameworks are designed from day one for robots interacting with the physical world. The founder argues 'Picking (grasping) is the Coding of embodied intelligence', using physical-world 'atomic tasks' as the entry point for embodied model scaling. It uses 'three-way data fusion' (internet semantics, autonomous-driving physical rules, robot manipulation data) in mixed training; after the Atomix merger it forms a 'scenario-type Scaling' data flywheel — massive real Picking data generated daily on the scenario side feeds model iteration, and stronger models in turn empower more warehouses and more robot types.

Commercialization（5）

• ◆Completed the equity-based merger with logistics robot company Atomix in June 2026; Atomix was spun out of Megvii by Tang Wenbin in July 2024, is a logistics robot heavyweight with nearly RMB 1B in annual revenue across over 20 countries and 500+ projects, and its real data directly fuels model training
• ◆Two revenue streams: infrastructure (zero-shot foundation-model MaaS billed by token) plus embodied application revenue (DFOL online post-training, OS deployment)
• ◆Customer segmentation: serves body makers and developers with the foundation model (DM0.5) + post-training services + OS to make third-party bodies smarter
• ◆Global layout: Hong Kong global HQ, Singapore as Atomix's global commercial HQ, Japan/Korea subsidiaries (close to Fast Retailing and other key customers), US office opened June 2026; a '100-person global recruitment plan' launched in parallel
• ◆Landing scenarios: warehouse logistics sorting (in synergy with Atomix), plus grasping and sorting of unstructured items such as clothing and soft packages

Position & Strengths（5）

• ◆A benchmark of the 'big-tech lineage' startup: Megvii's decade of vision algorithms, engineering and industry know-how transfers directly to embodied intelligence
• ◆Originator of the 'model × scenario' super-merger: the first in China's embodied track, closing the loop of brain, body and data ahead of peers
• ◆A top investor lineup: Alibaba + NIO + Zhipu + SenseTime + Stepfun + Huaqin Technology + SAIC Hengxu — strong endorsement from industrial capital (e-commerce + NEV + manufacturing + LLMs)
• ◆Open-source ecosystem: Dexbotic + RoboChallenge building community, co-building evaluation standards with Hugging Face
• ◆Originator of the 'Embodied Native' concept: its MMLA fusion-coexistence architecture became a mainstream consensus trend at WAIC 2026