近日，科大讯飞成立26周年之际，董事长刘庆峰发表内部讲话。 他坦言，相比使用英伟达等海外算力，基于国产算力开展大模型训练和算法创新存在诸多挑战，需要增 加移植适配、性能调优、对比验证等额外步骤和额外算力。"基于国产算力的所有新算法的研发会慢三 个月，在今天这个你追我赶的时代，是不得了的差别。" 但他同时也提到，科大讯飞要坚定不移地走自主可控之路，基于国产算力的新算法研发，会慢一些，但 不会矮。他给出了两个坚持自主可控的理由：一是因为有这个实力，能做，二是为了不被卡脖子，必须 做。 刘庆峰透露，在高考和世博会等场合表现不俗的星火X1，将在7月发布升级版。 在消费端，截至目前，学习机、办公本、翻译机、录音笔蝉联2025年"618"京东和天猫平台多个品类销 售冠军，都跑赢了大盘，全渠道收入同比增长25%，全周期预计增长超40%；软件上，听见SaaS当前总 用户数9000万，1-5月增长同比增加了54%，5G新通话产品当前总用户数706万，同比增加121%。 海外业务增长态势更为强劲。相比2024年，海外开发者数从38.6万增加到50.9万，同比增长31.9%；海 外AI营销业务营收，1-5月同比增长25倍；海外C端 ...

Core Insights - The chairman of iFlytek, Liu Qingfeng, emphasized the concept of "root dividend," focusing on self-control, integration of software and hardware, and deep industry engagement to drive innovation in large model technology and achieve true industrial advantages [1][2] - iFlytek's diverse applications showcase its advancements, including the X1 medical model reaching the level of chief physicians in top hospitals and the X1 legal model excelling in legal recommendations [1][3] - The company reported significant growth across its business segments, with a 25% year-on-year increase in revenue from products like learning machines and translation devices, and a 54% increase in users for its "Tingjian SaaS" software [1][3] B-end Developments - iFlytek has secured leading clients in the B-end sector, collaborating with major companies like PetroChina to develop the Kunlun large model and working with the State Energy Group on an intelligent evaluation assistant [2] G-end Applications - The AI-assisted case handling system for procuratorates has been launched for trial use, indicating progress in G-end applications [3] - The overseas market has shown rapid growth, with a 25-fold increase in AI marketing revenue and a threefold increase in C-end hardware sales from January to May [3] Technological Innovations - Liu Qingfeng highlighted three major technological innovations in iFlytek's large models, including a unified model training and reasoning method, reinforcement learning with fine-grained feedback, and a self-controlled reinforcement learning algorithm package [3] - The company is committed to training models based on domestic computing power, facing higher technical barriers and longer R&D cycles compared to overseas computing power, which is seen as essential for true self-control [3] Strategic Vision - Over its 26 years, iFlytek has transitioned from speech synthesis to cognitive intelligence and large models, with Liu Qingfeng stating the company aims to lead in foundational model technology while addressing real industry needs to enhance productivity and creativity [3]