On February 10, Dexmal held its first Know-how Open Day below the theme “Embodied Native,” marking the corporate’s first main public look since its founding.
Tang Wenbin, Co-founder and CEO of Dexmal, said: “2026 will not be the primary yr of embodied intelligence — it’s the first yr of ‘Embodied Native.’ What we’d like isn’t just a big mannequin that may run on robots, however a brand new AI paradigm whose intelligence and formation mechanisms are essentially rooted in bodily interplay.”
He defined that Embodied Native represents a brand new basis for large-scale, real-world deployment of robotics. The idea emphasizes ranging from scratch in defining information, coaching fashions, and deploying purposes primarily based on actual robotic situations — unifying basic interplay, producing generalized actions, and enabling robots to combine natively with real-world environments.
On the occasion, Dexmal launched DM0, described because the world’s first embodied-native basis mannequin educated fully from scratch.
DM0 incorporates three core options:
- Multi-source information pretraining
- Multi-task, cross-robot pretraining
- A spatial reasoning chain-of-thought mechanism
The mannequin’s most notable breakthrough lies in what the corporate calls its “intelligence density.” The two.4B parameter model ranked first globally on the authoritative real-world robotics benchmark RoboChallenge, securing prime positions in each single-task and multi-task classes.
Zhou Erjin, Accomplice at Dexmal, emphasised: “Reaching twin No.1 rankings with simply 2.4B parameters proves that parameter dimension will not be the barrier — native design is.”
Throughout pretraining, DM0 systematically blended three core activity classes — manipulation, navigation, and full-body management — throughout eight distinct robotic configurations, together with UR and Franka techniques. This pressured the mannequin to be taught generalized bodily operation logic reasonably than memorizing hardware-specific parameters, considerably enhancing cross-platform generalization.
DM0 additionally introduces a “spatial reasoning chain,” linking environmental notion, activity understanding, movement planning, and fine-grained execution right into a closed loop. This permits robots to finish advanced, multi-step actions requiring spatial reasoning — corresponding to taking a photograph after which sending an instruction.
To advertise technological democratization, Dexmal introduced that the two.4B model of DM0 is totally open-sourced, together with mannequin weights, code, and parameters and inference code for all 30 RoboChallenge duties.
Along with DM0, the corporate launched:
- Dexbotic 2.0, the world’s first embodied-native growth framework, that includes a modular structure that decouples the imaginative and prescient encoder, giant language mannequin, and motion skilled modules. This “Lego-style” design permits builders to independently improve or substitute parts throughout notion, cognition, and management layers, enabling fast adaptation to numerous {hardware} and utility situations.
- DFOL, a mass-production-oriented embodied-native workflow, designed round a “hardware-general + model-intelligent” paradigm to handle the longstanding trade-off between effectivity and adaptability in robotics industrialization.
By its Open Day, Dexmal clearly articulated its technical roadmap and enterprise philosophy: reducing growth boundaries by means of foundational native innovation, fostering ecosystem collaboration by means of open-source initiatives, and accelerating the transition of embodied intelligence from spectacular demos to scalable, real-world productiveness.
Supply: IPOzaozhidao
Elevate your perspective with NextTech Information, the place innovation meets perception.
Uncover the most recent breakthroughs, get unique updates, and join with a world community of future-focused thinkers.
Unlock tomorrow’s tendencies in the present day: learn extra, subscribe to our publication, and grow to be a part of the NextTech group at NextTech-news.com
