Dexmal Launches DM0.5 Embodied AI Model, DexOS and MaaS Platform

Dexmal has unveiled a comprehensive embodied AI ecosystem featuring its DM0.5 universal foundation model, DexOS operating system, MaaS (Model-as-a-Service) platform, DFOL 2.0 reinforcement learning framework, and the Apex universal robot. The launch follows the company's three-stage strategy to bridge the gap between advanced AI models and practical industrial deployment.

The new DM0.5 foundation model contains 4 billion parameters trained on more than 150,000 hours of multi-source data. Compared with its predecessor, the model delivers 25% higher inference efficiency, 50 ms inference latency, improved zero-shot and few-shot learning performance, and a leading score on the RoboChallenge Table30 V2 benchmark. The model also reduces downstream fine-tuning costs by approximately 60%.

Tang Wenbin, Chief Executive Officer of Dexmal, stated:

"Embodied AI has reached a critical stage where scaling laws are demonstrating their value. The remaining challenge is transforming powerful foundation models into real-world productivity through engineering integration. Our latest platform is designed to accelerate that transition across multiple robotic applications."

Supporting the foundation model are three software platforms designed for deployment at scale. DexOS separates AI models from robot hardware, enabling developers to control different robot platforms using a unified Python interface. DFOL 2.0 leverages a world model to reduce physical robot training requirements, while the MaaS platform simplifies cloud-based model training, deployment, and zero-shot inference through a pay-per-use service.

Dexmal also introduced the Apex universal robot featuring modular end-effectors, dual LiDAR-based 360-degree perception, millimeter positioning accuracy, a 3 kg payload per arm, and hot-swappable batteries for continuous operation.

The company is collaborating with robot manufacturers including Tiangong, Huaqin, and Shihe, alongside semiconductor partners such as Pingtouge, Moore Threads, and Cambricon, to expand China's embodied AI ecosystem.