Dexmal, formerly known as Yuanli Lingji, unveiled a comprehensive suite of embodied artificial intelligence products at its first Action developer conference. The goal of these developments is to address a fundamental industry problem—scaling foundational models onto diverse robotic hardware and in real-world conditions.
The centerpiece of the presentation was the embodied foundational model DM0.5. This model doubled its parameters to 4 billion and was trained on an extensive dataset comprising 150,000 hours of information. This volume includes 50,000 hours of data collected from real robots, annotated with key hand points in three dimensions with millimeter precision, as well as 100,000 hours of first-person human-controlled data.
DM0.5 features three architectural innovations. First, it offers native support for context memory up to 60 seconds, implemented via a pre-training abstraction layer. Second, a two-stage reasoning system has been introduced, which separates task planning from motion generation, utilizing counterfactual learning to ensure true instruction comprehension. Third, action alignment through limited dynamic programming is used to normalize trajectories across various execution speeds.
DexOS is positioned as a robotics operating system analogous to Android. It solves the N x M compatibility issue, where models and hardware require individual adaptation. By providing a standardized operating system with an open interface protocol, ECP, DexOS allows model developers to integrate once and then deploy across numerous different robotic platforms. The system unifies interaction with various sensors, actuators, and hardware configurations through a single API.
The DexDev platform integrates three functionalities: DFOL 2.0, which uses world-model-based reinforcement learning and reduces the need for real robot training data by 60%, as well as overall costs by 40% through simulation results in the DW0.5 world model; DexOS for cross-platform deployment; and MaaS (Model as a Service) for embodied AI inference on a pay-per-use basis. Dexmal demonstrated that the platform achieves an inference latency of 50 milliseconds and is capable of performing post-training on a single consumer RTX 4090 graphics card.
Furthermore, the company introduced two new robotic hardware platforms: Apex, designed for industrial applications, and Ferrata, focused on logistics, in addition to the existing dual-arm Mint platform. Founder Tang Wenbin expressed confidence that the path to mature embodied AI lies through developer ecosystems, not closed vertical integration. He noted that the true quality of a model is judged by whether developers choose it over open-source alternatives, stating that DM0.5 is Dexmal's first model that genuinely surpasses available open-source options.
