Dexmal was founded in March 2025. It is a scientific and technological innovation company oriented towards general embodied intelligence, with the mission of building intelligent, …
Apex universal robotUniversal embodied model DM0.5
Dexmal was founded in March 2025. It is a scientific and technological innovation company oriented towards general embodied intelligence, with the mission of building intelligent, useful and trustworthy robots. The company's core founding team comes from Megvii Technology, which possesses four major capabilities: large-scale basic model training, AI technology breakthroughs, large-scale implementation of AI robots, and high-reliability hardware mass production. It is a composite R&D team that is extremely rare in the industry. In terms of models, infrastructure, and data flywheels, they are the most typical representatives of domestic embodied large models.
02
MAIN PRODUCTS
Key Products
02 ITEMS
01
Apex universal robot
Dexmal Apex is an “embodied native” universal robot. By deeply integrating the underlying hardware with the embodied large model, the single limitation of traditional robot hardware is broken, and the original design concept and modular configuration are embodied, allowing the hardware to have the same freedom of combination and evolution capabilities as software: rich fine operation advantages: ultra-large torque, ultra-fast pace. Industrial grade and stable, built for mass production and high-intensity operations.
02
Universal embodied model DM0.5
DM0.5 is a universal embodied basic model for the open world. Compared with the previous generation DM0, DM0.5 has achieved comprehensive upgrades in four aspects: training data scale, data quality, architectural innovation, and open world generalization capabilities. The system capabilities have been upgraded from "executing fixed instructions and repeating a single action" to "understanding intentions, multi-dimensional continuous generalization, and long-term reliable operation." This model has surpassed the current optimal level in four major public evaluations of real machines and simulations. A single model can simultaneously support tasks such as navigation, grasping, and full-body control.
