![[9be78505-5564-425c-8c0d-d92a9c84d5d9.mp3]]

阶跃星辰开源3210亿参数MoE大模型Step 3

简报：

• 阶跃星辰（StepFun）宣布正式开源其最新一代基础大模型Step 3，该模型专为追求性能与成本均衡的企业和开发者设计。
• Step 3采用MoE架构，总参数量达3210亿，激活参数量380亿，具备强大的视觉感知和复杂推理能力。
• 模型通过自研MFA注意力和AFD优化，大幅提升了推理效率，并可在8×48GB显卡上完成大吞吐量推理。
• Step 3在MMMU、MathVision等多个权威评测集上表现领先，能够处理复杂的图文理解和逻辑推理任务。
• 模型及配套的StepMesh通信库已在Github、Hugging Face和魔搭ModelScope平台开源，API也已上线其开放平台。

相关链接：

• Github: https://github.com/stepfun-ai/Step3
• Hugging Face: https://huggingface.co/stepfun-ai/step3
• 魔搭ModelScope: https://www.modelscope.cn/models/stepfun-ai/step3

StepFun Open Sources 321 Billion Parameter MoE Large Model Step 3

Briefing:

• StepFun has officially announced the open-sourcing of its latest generation foundational large model, Step 3, which is designed for enterprises and developers seeking a balance between performance and cost.
• Step 3 adopts an MoE architecture with a total of 321 billion parameters and 38 billion active parameters, boasting powerful visual perception and complex reasoning capabilities. 🤖
• The model significantly enhances inference efficiency through self-developed MFA attention and AFD optimization, enabling high-throughput inference on 8×48GB GPUs.
• Step 3 demonstrates leading performance on multiple authoritative evaluation benchmarks such as MMMU and MathVision, capable of handling complex image-text understanding and logical reasoning tasks. 🌟
• The model and its accompanying StepMesh communication library have been open-sourced on GitHub, Hugging Face, and ModelScope platforms. The API is also available on its open platform. 🚀

Related Links:

• Github: https://github.com/stepfun-ai/Step3
• Hugging Face: https://huggingface.co/stepfun-ai/step3
• ModelScope: https://www.modelscope.cn/models/stepfun-ai/step3

火山引擎公布豆包大模型亮眼成绩，发布AI新品矩阵

简报：

• 字节跳动旗下火山引擎于7月31日宣布，豆包大模型日均tokens使用量已超16.4万亿，较去年5月增长137倍。
• 根据IDC报告，豆包大模型在中国公有云大模型服务调用量中以46.4%的市场份额位居第一。
• 发布了多款新产品，包括豆包・图像编辑模型3.0、豆包・同声传译模型2.0以及数字员工交互平台HiAgent。
• 新发布的豆包・同声传译模型2.0将语音延迟降至2-3秒，豆包・图像编辑模型3.0支持通过自然语言进行图像处理。

相关链接：

• 字节豆包大模型日均调用量激增 137 倍，推出多项新产品

VolcEngine Unveils Impressive Performance of Doubao Large Model, Releases New AI Product Matrix

Brief:

• ByteDance's VolcEngine announced on July 31st that the daily token usage of its Doubao large model has exceeded 16.4 trillion, a 137-fold increase compared to May last year. 🚀
• According to an IDC report, the Doubao large model ranks first in China's public cloud large model service call volume, holding a 46.4% market share. 🥇
• Several new products were released, including Doubao Image Editing Model 3.0, Doubao Simultaneous Interpretation Model 2.0, and the digital employee interaction platform, HiAgent.
• The newly released Doubao Simultaneous Interpretation Model 2.0 reduces voice latency to 2-3 seconds, while Doubao Image Editing Model 3.0 supports image processing via natural language. ✨

Related Links:

• 字节豆包大模型日均调用量激增 137 倍，推出多项新产品

苹果季度收入大增，宣布加码AI投资

简报：

• 苹果公司公布最新财报，总收入增长10%，创下自2021年12月以来的最大季度增幅。
• CEO蒂姆·库克表示，公司将“大幅增加”其在人工智能（AI）领域的投资。
• 库克称，苹果对通过并购来加速其发展路线图持开放态度。
• 在中国市场，苹果销售额同比增长4%，达到153.7亿美元，部分得益于当地的设备补贴政策。

相关链接：

• 苹果季度收入增幅创2021年年底以来最大，将“大幅增加”AI投资

Apple's Quarterly Revenue Surges, Announces Increased AI Investment

Briefing:

• Apple Inc. announced its latest financial results, with total revenue growing by 10%, marking its largest quarterly increase since December 2021. 📈
• CEO Tim Cook stated that the company will "significantly increase" its investment in Artificial Intelligence (AI). 🧠
• Cook also noted that Apple is open to accelerating its development roadmap through mergers and acquisitions (M&A).
• In the Chinese market, Apple's sales grew by 4% year-over-year, reaching $15.37 billion, partly benefiting from local device subsidy policies. 🇨🇳

Related Link:

• Apple's Quarterly Revenue Growth Highest Since End of 2021, Will 'Significantly Increase' AI Investment

苹果首款折叠iPhone或于2026年发布，定价近2000美元

简报：

• 根据摩根大通分析师报告，苹果预计将在2026年9月推出首款折叠iPhone，或命名为“iPhone 18 Fold”。
• 预计定价高达1999美元，可能成为苹果手机史上最贵的机型之一。
• 新品将采用“书本式”内折设计，配备由三星显示器提供的7.8吋无明显折痕屏幕。
• 功能上预计将取消Face ID，并回归实体的Touch ID指纹辨识。

相关链接：

• 蘋果摺疊機來了！「iPhone 18 Fold」預計2026登場 定價刷新高
• 折叠iPhone预计2026年登场，定价曝光

Apple's First Foldable iPhone May Launch in 2026, Priced Near $2000

Brief:

• According to a JPMorgan analyst report, Apple is expected to launch its first foldable iPhone in September 2026, potentially named "iPhone 18 Fold."
• It is anticipated to be priced as high as $1999, potentially making it one of the most expensive iPhone models in Apple's history. 🤑
• The new device will feature a "book-style" inward-folding design, equipped with a 7.8-inch display from Samsung Display, expected to have no noticeable crease. 📱
• Functionally, it is expected to drop Face ID and revert to physical Touch ID fingerprint recognition. 👍

Related links:

• 蘋果摺疊機來了！「iPhone 18 Fold」預計2026登場 定價刷新高
• 折叠iPhone预计2026年登场，定价曝光

Meta公布巨额AI投资计划，2025年资本支出高达720亿美元

简报：

• Meta在其第二季度财报中透露，为全面推进其人工智能宏图，计划将2025年的资本支出（包括融资租赁本金支付）大幅提升至660亿至720亿美元之间。
• 该投资将主要用于建设数据中心和服务器等AI基础设施，以增加算力，满足人工智能研发和业务运营的需求，并预计在2026年继续大幅增加投资。
• 为支持其AI愿景，Meta已公布两个AI“泰坦集群”项目：“普罗米修斯”（Prometheus）和“海伯利昂”（Hyperion）。
• 除了硬件设施，Meta的第二大增长驱动力是员工薪酬，公司正斥巨资为新成立的Superintelligence Labs招募顶尖AI人才。

相关链接：

• 相关报道
• 扎克伯格下血本：Meta 计划今年砸 720 亿美元加码 AI 基础设施

Meta Unveils Massive AI Investment Plan, 2025 Capex to Reach $72 Billion 🤖

Briefing:

• Meta revealed in its Q2 earnings report that to fully advance its artificial intelligence ambitions, it plans to significantly increase its capital expenditures (including principal payments for finance leases) for 2025 to between $66 billion and$72 billion. 🚀
• This investment will primarily be allocated to building AI infrastructure such as data centers and servers to boost computing power and meet the demands of AI research & development and business operations, with further substantial increases expected in 2026.
• To support its AI vision, Meta has announced two AI 'Titan Cluster' projects: 'Prometheus' and 'Hyperion'.
• Beyond hardware facilities, Meta's second largest growth driver is employee compensation. The company is spending heavily to recruit top-tier AI talent for its newly established Superintelligence Labs. 💰

Related Links:

• Related Report
• Zuckerberg's Big Bet: Meta Plans to Invest $72 Billion This Year in AI Infrastructure

GPT-5测试版模型现身Mac客户端，主打推理与自动化

简报：

• 开发者在OpenAI的Mac客户端（ChatGPT Agent应用）中发现了两个可能正在测试的新模型：GPT-5-Auto与GPT-5-Reasoning，暗示其下一代AI模型已进入内部测试阶段。
• 发现的代码片段显示，名为“gpt-5-reasoning-alpha-2025-07-13”的模型配置包含高强度推理设置。
• GPT-5-Reasoning被认为是o系列推理模型的延续，专注于复杂逻辑与多步推理；GPT-5-Auto则被认为是一个自动化智能体，可集成外部服务自主执行任务。
• 这些线索表明，结合了多模态与推理能力的GPT-5可能最早于2025年夏季正式发布。

相关链接：

• https://www.aibase.com/zh/news/20132
• https://www.aibase.com/zh/news/20115

GPT-5 Beta Models Emerge in Mac Client, Highlighting Reasoning and Automation 🧠

Brief:

• Developers have discovered two potentially new models under testing in OpenAI's Mac client (ChatGPT Agent application): GPT-5-Auto and GPT-5-Reasoning, suggesting its next-generation AI model has entered internal testing. 🚀
• Discovered code snippets reveal that the model named "gpt-5-reasoning-alpha-2025-07-13" includes high-intensity reasoning configurations.
• GPT-5-Reasoning is believed to be a continuation of the 'o' series reasoning models, focusing on complex logic and multi-step inference; GPT-5-Auto is thought to be an autonomous agent capable of integrating external services to execute tasks independently.
• These clues suggest that GPT-5, combining multimodal and reasoning capabilities, could be officially released as early as summer 2025. ✨

Related Links:

• https://www.aibase.com/zh/news/20132
• https://www.aibase.com/zh/news/20115

阿里开源网络智能体WebAgent，多项评测创佳绩

简报：

• 阿里云通义实验室近日宣布开源其自主搜索AI智能体项目WebAgent，该项目旨在模拟人类在网络环境中进行信息检索和多步推理。
• 项目包含两大核心组件：采用“形式化驱动”数据合成方法的WebShaper，以及作为智能体“大脑”的大语言模型WebSailor。
• 在权威基准测试中，WebAgent在GAIA评测上获得60.19分，在WebWalkerQA评测上获得52.2分。
• 其WebSailor-72B模型在BrowseComp评测集上的表现超越了多个闭源模型，登顶开源网络智能体榜单。
• 项目代码、数据集及模型已在GitHub、Hugging Face和ModelScope等平台开放，并在GitHub获得广泛关注。

相关链接：

• 阿里开源WebAgent项目WebShaper GAIA评测超过Claude4-Sonnet
• 阿里通义千问发布 Qwen3-Coder-Flash 模型

Alibaba Open-Sources WebAgent, Achieving Excellent Results in Multiple Benchmarks

Briefing:

• Alibaba Cloud's Tongyi Lab recently announced the open-sourcing of its autonomous search AI agent project, WebAgent, which aims to simulate human information retrieval and multi-step reasoning in a web environment.
• The project comprises two core components: WebShaper, which uses a "formal-driven" data synthesis method, and WebSailor, a large language model serving as the agent's "brain."
• In authoritative benchmark tests, WebAgent scored 60.19 points on the GAIA evaluation and 52.2 points on the WebWalkerQA evaluation. 🏆
• Its WebSailor-72B model outperformed several closed-source models on the BrowseComp benchmark, topping the open-source web agent leaderboard. 🚀
• The project code, datasets, and models are now available on platforms like GitHub, Hugging Face, and ModelScope, and have gained significant attention on GitHub. 🌐

Related Links:

• Alibaba Open-Sources WebAgent Project WebShaper Surpasses Claude4-Sonnet in GAIA Evaluation
• Alibaba Tongyi Qianwen Releases Qwen3-Coder-Flash Model