发布全球首个具身原生大模型DM0，原力灵机唐文斌：今年将是「具身原生」元年

Core Viewpoint - The article emphasizes the transition of embodied intelligence from impressive demonstrations to widespread, scalable productivity through innovative technology and open-source collaboration [7]. Group 1: Company Overview - Yuanli Lingji held its first "Embodied Native - Yuanli Lingji Technology Open Day" on February 10, marking its first major public appearance since establishment [3]. - CEO Tang Wenbin stated that 2026 will not be the year of embodied intelligence but rather the year of "embodied native," focusing on a new AI paradigm rooted in physical interaction [3]. Group 2: Technological Innovations - Yuanli Lingji introduced the world's first zero-start training embodied native large model DM0, characterized by multi-source data pre-training, multi-task and cross-model pre-training, and spatial reasoning thinking chains [4]. - DM0 achieved a significant breakthrough with its 2.4 billion parameter version, ranking first in both single-task and multi-task categories on the RoboChallenge benchmark, currently holding the top position globally [4][6]. Group 3: Model Features and Capabilities - The model's design emphasizes that parameter quantity is not a barrier; rather, the native design is crucial for performance [6]. - DM0 integrates core tasks such as operation, navigation, and full-body control during pre-training, covering eight different robot hardware configurations, which enhances its generalization capabilities across models [6]. - The construction of a "spatial reasoning thinking chain" allows the robot to perform complex actions requiring multi-step spatial reasoning, such as "taking a photo and sending instructions" [6]. Group 4: Open Source and Development Framework - To democratize technology, Yuanli Lingji announced that the 2.4 billion parameter version of DM0 will be fully open-sourced, including code, models, and parameters for all 30 tasks in RoboChallenge [6]. - The company also launched the world's first embodied native development framework, Dexbotic 2.0, which features a modular architecture allowing developers to independently upgrade or replace components [6]. - The DFOL workflow aims to solve the industry dilemma of balancing efficiency and flexibility through a model that combines hardware universality with intelligent models [6].