On January 13, 2026, Moore Threads and the Beijing Academy of Artificial Intelligence (BAAI) jointly made an announcement that they had successfully wrapped up the full - process training of the embodied brain model, RoboBrain 2.5. This achievement was made possible by utilizing the domestically - manufactured MTT S5000 thousand - card intelligent computing cluster and the FlagOS - Robo framework. In the global context of AI development, computing resources play a crucial role. This training serves as a solid proof of the usability and efficiency of domestically - produced computing clusters when it comes to training embodied AI large models. It stands as a significant milestone, indicating that China's AI infrastructure is now well - equipped to handle complex multimodal tasks. RoboBrain 2.5 has been endowed with brand - new capabilities. It can now understand and reason about the temporal value of actions as well as three - dimensional spatial structures. This enhancement has led to a remarkable improvement in the success rate of downstream task execution. When comparing the models trained on the MTT S5000 with those trained using internationally mainstream GPUs, the MTT S5000 - trained models show a high level of consistency across multiple key metrics. In some tasks, they even outperform their GPU - trained counterparts. Moreover, the training curves of these two types of models are highly overlapping, with a relative error of less than 0.62%. Another impressive feat of the MTT S5000 thousand - card intelligent computing cluster is its scaling ability. When expanding from 64 cards to 1024 cards, it has achieved over 90% linear scaling efficiency, proving that it is capable of supporting training on a ten - thousand - card scale. This collaboration between Moore Threads and BAAI offers the industry a replicable and scalable "domestic computing power training paradigm." It is set to speed up the transition of embodied AI from the realm of laboratory research to real - world industrial implementation.