![[57877f4c-aa96-4c9f-b179-6c5a03494d51.mp3]]

智谱AI发布开源多模态视觉语言大模型GLM-4.1V-Thinking，性能对标国际顶尖产品

简报：

• 智谱AI正式开源通用视觉语言模型GLM-4.1V-Thinking，支持图像、视频及文档等多模态输入，采用创新思维链推理机制与课程采样强化学习策略，显著提升跨模态因果推理能力与稳定性。
• GLM-4.1V-Thinking轻量化实现（9B参数），在28项权威多模态评测中23项成绩达10B级模型最佳，18项持平或超越72B参数级的Qwen-2.5-VL，关键任务表现与OpenAI等全球顶尖模型比肩甚至超越。
• 新模型支持64K上下文与4K图像处理，具备中英文双语能力，广泛应用于长视频理解、图像问答、学科解题、GUI操作等领域，免费商用授权，单张3090显卡可部署。
• 模型权重已在Hugging Face与魔搭社区同步开源，进一步推动中国AI产业国际影响力提升，与OpenAI、Google等国际巨头正面竞争。

相关链接：

• AIbase新闻报道
• IT之家详细介绍

Zhipu AI Releases Open-Source Multimodal Vision-Language Model GLM-4.1V-Thinking, Performance Benchmarks International Top Products 🤔

Briefing:

• Zhipu AI officially open-sources its general vision-language model GLM-4.1V-Thinking, supporting multimodal inputs such as images, videos, and documents. It adopts an innovative Chain-of-Thought (CoT) reasoning mechanism and curriculum sampling reinforcement learning strategy, significantly enhancing cross-modal causal reasoning capabilities and stability.
• GLM-4.1V-Thinking achieves a lightweight implementation (9B parameters). In 28 authoritative multimodal evaluations, it achieved the best results among 10B-class models in 23 tasks and equaled or surpassed the 72B-parameter Qwen-2.5-VL in 18 tasks. Its performance on key tasks is comparable to, or even surpasses, global leading models like OpenAI's. 🚀
• The new model supports 64K context and 4K image processing, possesses bilingual (Chinese and English) capabilities, and is broadly applicable in areas such as long video understanding, image Q&A, subject problem-solving, and GUI operations. It's available with free commercial licensing and can be deployed on a single 3090 graphics card.
• Model weights are simultaneously open-sourced on Hugging Face and ModelScope communities, further boosting the international influence of China's AI industry and directly competing with international giants like OpenAI and Google. 🌍

Related Links:

• AIbase News Report
• IT Home Detailed Introduction

富士康推出台湾首款自研AI推理大模型“FoxBrain”，商标进入实质审查待审阶段

简报：

• 富士康（鸿海精密工业股份有限公司）正式推出首款自研AI推理大模型“FoxBrain”，并已向国家知识产权局提交相关商标注册申请，当前状态为“等待实质审查”。
• “FoxBrain”由鸿海研究院开发，涵盖数据分析、数学推理和代码生成等多项功能，是台湾首个该类别AI模型，尤其针对繁体中文进行了优化。
• 初始版本基于Meta Llama 3.1，采用120块英伟达H100 GPU训练一个月打造，但与部分对手模型如DeepSeek相比性能略有差距。
• 联发科此前亦推出Llama-Breeze2系列AI模型，主打繁体中文处理及轻量化，显示台湾科技企业在AI领域布局加速。

相关链接：

• Foxconn unveils ‘FoxBrain,’ built on Nvidia GPUs to boost AI efforts - Yahoo Finance
• IT之家相关报道
• Aibase相关新闻

Foxconn Unveils "FoxBrain," Taiwan's First Self-Developed AI Inference Large Model; Trademark Awaiting Substantive Examination

Briefing:

• Foxconn (Hon Hai Precision Industry Co., Ltd.) has officially launched its first self-developed AI inference large model, "FoxBrain." The company has submitted a trademark registration application to the National Intellectual Property Administration, which is currently in the "awaiting substantive examination" status. 🧠
• "FoxBrain," developed by Hon Hai Research Institute, offers multiple functionalities including data analysis, mathematical reasoning, and code generation. It is Taiwan's first AI model of its kind, specifically optimized for Traditional Chinese.
• The initial version is based on Meta Llama 3.1 and was trained for a month using 120 Nvidia H100 GPUs. However, its performance slightly lags behind some competitor models like DeepSeek. 🚀
• MediaTek also previously launched its Llama-Breeze2 series of AI models, focusing on Traditional Chinese processing and lightweight design, indicating an accelerated expansion of Taiwanese tech companies in the AI sector. 🇹🇼

Related Links:

• Foxconn unveils ‘FoxBrain,’ built on Nvidia GPUs to boost AI efforts - Yahoo Finance
• IT之家相关报道
• Aibase相关新闻

百度推出MuseSteamer模型，开启中文AI音视频生成新时代

简报：

• 2025年7月2日，百度商业研发团队发布全球首个实现中文音视频一体化生成的视频模型MuseSteamer，并同步上线“绘想”创作平台。
• MuseSteamer支持用户仅需上传一张图片，即可生成10秒1080p电影级、有声动态视频，画面、音效和人声台词协同生成，微表情和运镜效果达到专业影视水平。
• 在权威榜单VBench I2V中，MuseSteamer以89.38%总分位列全球首位，模型家族覆盖Turbo、Lite、Pro及全有声版，多版本将于今年8月陆续开放，“绘想”平台现已开启 Turbo 限时免费公测。
• 此技术有望极大降低视频创作门槛，激发内容多样性，推动非专业用户参与专业级视听内容制作。

相关链接：

• IT之家：百度发布自研视频生成模型 MuseSteamer
• AIBASE新闻1
• AIBASE新闻2
• 百度“绘想”平台

Baidu Launches MuseSteamer Model, Ushering in a New Era of Chinese AI Audio-Video Generation

Briefing:

• On July 2, 2025, Baidu's Business R&D team unveiled MuseSteamer, the world's first video model capable of integrated Chinese audio-video generation, simultaneously launching the "Huixiang" creation platform. 🚀
• MuseSteamer allows users to generate 10-second, 1080p cinematic, dynamic videos with sound by simply uploading a single image. Visuals, sound effects, and human voice dialogue are collaboratively generated, with micro-expressions and camera movement effects reaching professional film and television standards. 🎬
• MuseSteamer secured the global top spot on the authoritative VBench I2V ranking with an overall score of 89.38%. The model family includes Turbo, Lite, Pro, and full-audio versions, with multiple versions set to be progressively released starting this August. The "Huixiang" platform has already commenced limited-time free public beta for the Turbo version.
• This technology is expected to significantly lower the barrier to video creation, stimulate content diversity, and empower non-professional users to participate in professional-grade audio-visual content production. ✨

Related Links:

• IT Home: Baidu Releases Self-Developed Video Generation Model MuseSteamer
• AIBASE News 1
• AIBASE News 2
• Baidu "Huixiang" Platform

OpenAI推出企业级AI定制与咨询服务，起步价达千万美元

简报：

• OpenAI已加大企业AI咨询业务力度，向组织单位提供定制化模型（如GPT-4o）、数据标注和应用开发服务，每位客户最低收费为1000万美元。
• 公司工程师将直接与客户协作，以开发专属的聊天机器人等AI应用，并考虑将部分数据标注任务外包以提升效率。
• 现有客户涵盖美国国防部、东南亚科技公司Grab等，OpenAI此举旨在强化企业级AI整合与推动智能化转型，与Palantir、Accenture等企业形成竞争。

相关链接：

• OpenAI启动高端企业AI咨询服务，客户最低收费1000万美元
• OpenAI加码AI咨询业务 1000万美元起步对标Palantir/埃森哲
• OpenAI推出企业AI定制与咨询，正式对标大型IT服务商

OpenAI Launches Enterprise AI Customization and Consulting Services, Starting at $10 Million 🚀

Brief:

• OpenAI has significantly scaled up its enterprise AI consulting business, offering organizations tailored models (such as GPT-4o), data labeling, and application development services, with a minimum fee of $10 million per client. 💰
• Company engineers will collaborate directly with clients to develop exclusive AI applications like chatbots, and are exploring outsourcing some data labeling tasks to enhance efficiency.
• Existing clients include the U.S. Department of Defense and Southeast Asian tech company Grab. This initiative by OpenAI aims to strengthen enterprise-level AI integration and drive intelligent transformation, positioning it in competition with firms like Palantir and Accenture. 🧠

Related Links:

• OpenAI Launches High-End Enterprise AI Consulting Services, Minimum Client Fee of $10 Million
• OpenAI Boosts AI Consulting Business, Starting at $10 Million to Rival Palantir/Accenture
• OpenAI Launches Enterprise AI Customization and Consulting, Officially Competing with Large IT Service Providers

微软推出MAI-DxO医疗AI，诊断准确率大幅超越资深专家

简报：

• 微软首席执行官纳德拉宣布推出医疗AI系统 MAI-DxO，采用“模型无关”设计，能适配不同厂商的语言模型以提升诊断性能。
• 系统能够模拟医生诊断流程，在官方公布的测试中，MAI-DxO诊断准确率高达85.5%，远超21名十年以上经验医生的19.9%平均水平。
• MAI-DxO创新性地引入虚拟医生团队协作机制，并支持五种适用于不同医疗场景的集成模式，显著提升诊断效率和成本效益。
• 微软同时发布了医疗序贯诊断基准SDBench，以304个真实医学案例检验AI及人类医生的序贯诊断能力，树立了医疗AI行业新标准。

相关链接：

• AIBase报道
• IT之家新闻

Microsoft Unveils MAI-DxO Medical AI, Dramatically Outperforming Senior Experts in Diagnostic Accuracy

Brief:

• Microsoft CEO Satya Nadella announced the launch of MAI-DxO, a medical AI system featuring a "model-agnostic" design that can adapt to language models from various vendors to enhance diagnostic performance. 🚀
• The system can simulate a doctor's diagnostic process. In officially released tests, MAI-DxO achieved a diagnostic accuracy of 85.5%, significantly surpassing the average level of 19.9% from 21 doctors with over ten years of experience. 🤯
• MAI-DxO innovatively introduces a virtual doctor team collaboration mechanism and supports five integration modes applicable to various medical scenarios, significantly improving diagnostic efficiency and cost-effectiveness.
• Microsoft also released SDBench, a medical sequential diagnosis benchmark, which tests the sequential diagnostic capabilities of AI and human doctors using 304 real medical cases, setting a new standard for the medical AI industry. 💡

Related Links:

• AIBase Report
• IT Home News