L3 高阶智驾专题暨 VLA 模型产业 自皮书及产业图谱发布 2025年9月4日(周四) 上海 - 国泰海通外滩金融广场 大模型发展如火如荼,将大模型进一步融合至智能驾驶中已成为产业共识, 而近年来政策正使得 L3 级智能驾驶落地商用渐成可能。在此背景下,视觉 - 酒言。动作趋刑(八)应云而生 VIA右望幼稚米们人米智融目的敕休计年 框架,将影响智能驾驶、具身智能产业格局与技术发展路线,并带来巨大的市 场和资本机遇。 13:30-13:40 领导发言致辞 陈忠义 - 国泰海通证券副总裁、研究与机构业务委员 会总裁、研究所党委书记、政策和产业研究院院长 吴珩 - 上汽集团金融事业部 总经理、上汽集团金控管 理有限公司 总经理 13:40-14:00 通往 L3 智能驾驶与具身智能之钥 -- 视觉 - 语言 - 动作模型(VLA) 朱峰 - 国泰海通政策和产业研究院科技首席分析师 14:00-14:30 主题演讲 袁玉记 -Momenta 解决方案总监 Momenta (北京初速度科技有限公司)是全球领先自 动驾驶公司,致力于通过深度学习和人工智能技术实 现可规模化的自动驾驶解决方案。公司基于数据驱动 的"一个飞 ...

Core Insights - The core focus of the articles is on Honor's significant advancements in edge AI voice technology, particularly through the acceptance of two research papers at the prestigious INTERSPEECH 2025 conference, highlighting the company's commitment to innovation in multilingual real-time voice recognition and translation [1][2]. Group 1: Research Achievements - Honor's two papers, in collaboration with Shanghai Jiao Tong University, address critical challenges in achieving high accuracy and responsiveness in translation experiences on mobile devices with limited computational resources [2][5]. - The research emphasizes the transition from academic findings to practical applications, showcasing a seamless integration of technology into real-world multilingual communication solutions [5]. Group 2: Technological Innovations - Honor's research project aims to deliver a cloud-comparable translation experience purely on-device, overcoming the dual challenges of low latency and high accuracy under resource constraints [6]. - The company introduced two innovative technical solutions: a novel attention mechanism enabling real-time speech recognition without waiting for complete sentences, and a speculative sampling inference module that enhances prediction speed while maintaining accuracy [6][7]. Group 3: Performance Metrics - The new technologies have demonstrated impressive results, reducing memory usage from 3-4GB to 800MB, achieving a 75% reduction in storage requirements, a 16% increase in translation accuracy, and a 38% improvement in inference speed [7]. - Honor has successfully developed the world's first edge voice large model, integrating six languages into a model with only 0.8 billion parameters, allowing offline processing and ensuring user privacy [7]. Group 4: Strategic Vision - Honor's breakthroughs in edge AI voice technology are a result of its long-term commitment to AI strategy, marked by consistent investment and a clear evolution path from AI experience implementation to the development of large models [9]. - The recent launch of the self-developed multi-modal perception large model, MagicGUI, with 70 billion parameters, further solidifies Honor's position in the AI landscape, achieving industry-leading capabilities [9].