Group 1 - The 2025 East Lake International Artificial Intelligence Summit Forum was held in China, focusing on accelerating the development of new productive forces and responding to the "Artificial Intelligence+" action plan [1] - The launch of the ZDTC 4.0 multimodal reasoning model marks a significant breakthrough in high-level semantic understanding and reasoning capabilities, transitioning from "pure text thinking" to "fine-grained multimodal semantic thinking" [1] - The Wuhan Artificial Intelligence Research Institute aims to leverage the ZDTC multimodal model to create internationally advanced AI application results and technology patents, contributing to regional industrial upgrades [1] Group 2 - The ZDTC Cloud is the first native collaborative multimodal large model cloud in China, providing comprehensive capabilities from computing power support to application implementation [2] - The platform aims to empower core business operations of enterprises by leveraging full-stack AI capabilities, focusing on creating benchmark cases in Hubei's advantageous industries [2] - The initiative is expected to contribute to the intelligent transformation of China's modern industrial system [2]

Core Insights - The launch of the ZDTC 4.0 multimodal reasoning model and the "ZDTC Cloud" platform aims to promote the industrialization of domestic large models, with 28 companies signing agreements to become ecosystem partners [2][5]. Group 1: Model Development - The ZDTC 4.0 model represents a significant milestone in the evolution of multimodal large models, transitioning from passive analysis to active reasoning capabilities [3]. - The model can integrate multimodal information and generate explainable intelligent decisions, showcasing its ability to respond in seconds to user queries, such as providing travel information based on a photo [3][4]. - Compared to the previous version, ZDTC 3.0, the 4.0 version demonstrates enhanced fine-grained multimodal interaction and reasoning abilities, allowing it to mimic human attention mechanisms [3][4]. Group 2: Applications and Impact - The ZDTC model has been applied in over 100 scenarios, including a smart service platform for orthopedic instruments that manages thousands of medical supplies [5]. - The "ZDTC Cloud" platform offers comprehensive support for enterprises, covering computing power services, model training, application development, and embodied intelligence [5]. - The model has been benchmarked against international standards in multimodal understanding and logical reasoning, achieving breakthroughs in tasks such as long video comprehension [4]. Group 3: Future Prospects - The company aims to create more benchmark cases in low-altitude economy, embodied intelligence, manufacturing, and healthcare, with a revenue target of 100 million yuan by 2025 [5]. - The establishment of partnerships with various organizations, including the "AI + Meteorology" innovation studio, indicates a strategic focus on collaborative development in AI technology [7].

9月19日，首款全栈国产化深度推理大模型"紫东太初"4.0在武汉发布，具备"边看、边识、边思"的类人多模态推理认知能力，特别是在带图思考多模态复杂 推理和工具调用的能力上，全面超过了GPT5。 2024年图灵奖得主理查德·萨顿表示，人类数据正在迅速接近极限，智能体将越来越多通过与世界交互而获得持续学习的能力，潜力远超以往。 中国科学院自动化研究所副总工程师、武汉人工智能研究院院长王金桥在接受第一财经记者采访时介绍，从3.0原生的多模态统一框架到4.0多模态细腻度的 复杂思考，紫东太初完成了国产大模型从"纯文本思考""简单操作带图思考"到"细粒度多模态语义思考"的三重跃迁。 以推理"赢得一场斯诺克需要几步"为例，"紫东太初"4.0通过输入的图片可先定位白球、粉球和红球的位置，再通过台上不同球的得分进行复杂的推理和数 学计算，进一步告诉大家他要做5杆斯洛克才能赢得这场球。 "紫东太初4.0从理解能力、模态推理、动态推理三大能力引入了类人的交叉注意力机制，使大模型能够像人一样，实现对世界的理解从全局到局部再到上下 文交互式的思考过程。"王金桥说，通过视觉注意力关注到某一个区域，然后进行平移、放大、旋转、增强等一系列 ...

9月19日，2025东湖国际人工智能高峰论坛在武汉举行。会上，紫东太初4.0多模态推理大模型正式发 布，该大模型由中国科学院自动化研究所与武汉人工智能研究院共同打造。 武汉人工智能研究院院长王金桥接受证券时报记者采访时表示，紫东太初4.0具备"边看、边识、边 思"的认知能力，在高阶语义理解与推理能力上实现重要突破，技术演进层面完成了从"纯文本思 考"、"简单操作带图思考"到"细粒度多模态语义思考"的三重跃迁，迈向通用多模态推理新阶段。 此外，现场共有28家企业签约成为"紫东太初"生态合作伙伴。"紫东太初"生态合作体系覆盖算力芯片、 具身智能、数据智能、行业应用等领域。 以武汉紫东太初为代表，武汉已崛起人工智能产业集群。此次论坛上公布的数据显示，近3年来，武汉 人工智能产业年均增长30%以上，2024年产业规模超过700亿元。 目前武汉已聚集人工智能相关企业1000余家，全市在用的10亿参数规模以上的大模型60余个，初步形成 覆盖基础层、技术层、应用层的人工智能全产业链。 近年来，武汉出台实施促进人工智能产业发展系列政策举措，聚焦智能芯片、智能终端、智能网联汽 车、智能装备四大细分赛道，推进智能穿戴、智能座舱、 ...

Core Insights - The 2025 East Lake International Artificial Intelligence Summit Forum was held in Wuhan, where the ZDTC 4.0 multimodal reasoning model was officially launched, developed by the Chinese Academy of Sciences and Wuhan Artificial Intelligence Research Institute [1] - The ZDTC 4.0 model demonstrates significant breakthroughs in high-level semantic understanding and reasoning capabilities, evolving from "pure text thinking" to "fine-grained multimodal semantic thinking" [1] - The model can achieve deep understanding of 180-minute long videos and provide precise answers in seconds, topping six datasets in long video reasoning and retrieval capabilities [1] Industry Developments - The ZDTC Cloud platform was launched as the first native collaborative cloud for multimodal large models in China, offering comprehensive capabilities from computing power support to application implementation [2] - Over the past three years, Wuhan's AI industry has grown by more than 30% annually, with the industry scale expected to exceed 70 billion yuan in 2024 [2] - Wuhan has gathered over 1,000 AI-related companies and more than 60 large models with over 1 billion parameters in use, forming a complete AI industry chain [2] Policy and Innovation - Wuhan has implemented a series of policies to promote AI industry development, focusing on smart chips, smart terminals, smart connected vehicles, and smart equipment [3] - The city is advancing product innovation and industrialization in areas such as smart wearables, smart cockpits, and humanoid robots [3]

为推动紫东太初4.0的技术优势转化为实际产业价值，"紫东太初云"平台同步发布。 9月19日，首款全栈国产化深度推理大模型"紫东太初"4.0在武汉发布，具备"边看、边识、边思"的类人多模态推理认知能力，特别是在带图思考多模态复杂 推理和工具调用的能力上，全面超过了GPT5。 2024年图灵奖得主理查德·萨顿表示，人类数据正在迅速接近极限，智能体将越来越多通过与世界交互而获得持续学习的能力，潜力远超以往。 中国科学院自动化研究所副总工程师、武汉人工智能研究院院长王金桥在接受第一财经记者采访时介绍，从3.0原生的多模态统一框架到4.0多模态细腻度的 复杂思考，紫东太初完成了国产大模型从"纯文本思考""简单操作带图思考"到"细粒度多模态语义思考"的三重跃迁。 以推理"赢得一场斯洛克需要几步"为例，"紫东太初"4.0通过输入的图片可先定位白球、粉球和红球的位置，再通过台上不同球的得分进行复杂的推理和数 学计算，进一步告诉大家他要做5杆斯洛克才能赢得这场球。 在视频多模态2AG应用上，"紫东太初"4.0能够实现180分钟长视频的深度理解，细腻度片段定位和内容自动分析、总结，在视频问答、细腻度检索和文档检 索等6项任务中达 ...