![[8562a533-0ae7-41f4-836d-523b9e68f1a2.mp3]]

字节跳动火山引擎“奇美拉”数字人平台开启封测，月底有望公测

简报：

• 火山引擎正在封闭测试新一代数字人平台“奇美拉”，平台由字节跳动旗下智能创作数字人团队打造。
• 奇美拉平台基于火山引擎AI大模型技术，提供数字人、图片换装、视频翻译等服务，目前采用定向邀请模式邀测，测试阶段免费，预计本月底启动公测。
• 平台正式上线后将以使用次数或视频生成时长计费，拓展数字人在金融、直播、营销等多场景应用。

相关链接：

• 36氪新闻
• AIbase相关报道

ByteDance VolcEngine "Chimera" Digital Human Platform Enters Closed Beta, Public Beta Expected by Month-End 🚀

Brief:

• VolcEngine is currently conducting closed beta tests for "Chimera," its new generation digital human platform. The platform is developed by ByteDance's intelligent creation digital human team.
• The Chimera platform leverages VolcEngine's large AI model technology, offering services such as digital human generation, virtual try-on for images, and video translation. It is currently operating on an invitation-only basis, with the testing phase being free. Public beta is anticipated to launch by the end of this month. 🗓️
• Upon official launch, the platform will charge based on usage frequency or video generation duration, aiming to expand the application of digital humans across various scenarios including finance, live streaming, and marketing. 💼

Related Links:

• 36Kr News
• AIbase Related Report

Meta高薪招揽AI精英，新组“超级智能实验室”抢占AI竞赛先机

简报：

• Meta近期成立“超级智能实验室”（Superintelligence Labs），大幅扩充AI团队以强化基础模型研发。
• 实验室已招募44名顶尖AI人才，其中40%曾在OpenAI工作，50%来自中国，70%为研究人员，75%拥有博士学位，团队兼具多元背景。
• Meta首席执行官扎克伯格转向AI领域，积极从OpenAI、苹果和谷歌DeepMind等公司高薪挖角，部分签约奖金高达1亿至2亿美元。
• 新员工多入职未满一个月，年薪范围在1000万至1亿美元，突出AI领域人才竞争加剧与Meta雄心。
• Meta此举将显著推动其在人工智能市场的布局，预示未来AI市场竞争升级。

相关链接：

• Meta大举扩张AI团队：顶尖人才纷纷加盟，薪资高达1亿

Meta Poaches Top AI Talent, Establishes "Superintelligence Labs" to Gain Edge in AI Race 💰

Brief:

• Meta recently established "Superintelligence Labs," significantly expanding its AI team to strengthen foundational model research and development.
• The lab has recruited 44 top AI talents: 40% previously worked at OpenAI, 50% are from China, 70% are researchers, and 75% hold PhDs, showcasing a diverse team background.
• Meta CEO Mark Zuckerberg has pivoted towards AI, actively poaching talent with high salaries from companies like OpenAI, Apple, and Google DeepMind, with some signing bonuses reportedly reaching $100 million to$200 million.
• Most new hires joined less than a month ago, with annual salaries ranging from $10 million to$100 million, underscoring the escalating talent competition in the AI field and Meta's ambition. 🚀
• This strategic move by Meta is expected to significantly advance its position in the artificial intelligence market, foreshadowing intensified AI market competition. 🧠

Related Link:

• Meta Expands AI Team Significantly: Top Talent Joins, Salaries Up to $100 Million

OpenAI宣布2025年底前部署超百万GPU，推动AI基础设施扩张

简报：

• OpenAI首席执行官萨姆·奥尔特曼近日表示，公司计划到2025年底前上线超过100万个GPU，体现其在人工智能领域的扩容雄心和“百倍扩容”技术愿景。
• 该目标是OpenAI全球AI基础设施战略的重要组成部分，Stargate项目成为核心，未来四年预计投资高达5000亿美元，建设全球最大AI训练集群，首期工程布局在得克萨斯州阿比林市。
• 项目首期投资方包括软银、OpenAI、甲骨文等，软银CEO孙正义任Stargate董事长，OpenAI负责日常运营；技术合作伙伴涵盖Arm、微软、英伟达、甲骨文等全球行业巨头。
• 业界普遍认为，此举将强化OpenAI未来在AI基础设施和技术竞争中的潜在优势，加速AI技术发展。

相关链接：

• OpenAI计划到2025年底启用100万个GPU
• OpenAI在阿比林建全球最大AI训练集群

OpenAI Announces Deployment of Over One Million GPUs by End of 2025, Driving AI Infrastructure Expansion

Brief:

• OpenAI CEO Sam Altman recently stated that the company plans to bring online over 1 million GPUs by the end of 2025, showcasing its ambitious scaling plans and "100x scaling" technological vision in the field of AI. 🚀
• This goal is a key component of OpenAI's global AI infrastructure strategy, with the Stargate project at its core, projected to involve an investment of up to $500 billion over the next four years to build the world's largest AI training cluster. The initial phase is planned for Abilene, Texas.
• Initial investors in the project's first phase include SoftBank, OpenAI, Oracle, and others. SoftBank CEO Masayoshi Son will serve as Stargate's chairman, with OpenAI responsible for daily operations. Technological partners encompass global industry giants such as Arm, Microsoft, NVIDIA, and Oracle. 🤖
• The industry widely believes this move will strengthen OpenAI's potential advantage in future AI infrastructure and technological competition, accelerating the development of AI technology. 💡

Related Links:

• OpenAI Plans to Activate 1 Million GPUs by End of 2025
• OpenAI to Build World's Largest AI Training Cluster in Abilene

台积电凭AI浪潮市值首破1万亿美元创历史新高

简报：

• 台积电市值近日在台湾股市首次突破1万亿美元，主要受益于强劲的人工智能（AI）需求和上调的业绩展望。
• 公司为苹果、英伟达等企业代工芯片，AI产业的扩张推动其股价自4月低点上涨近50%，2025年第二季度财报显示营收与净利双双刷新纪录。
• 台积电还上调了2023年全年营收增长预期至约30%，高盛和摩根大通分析师均看好台积电先进制程和AI客户支出的持续强劲。
• 市场普遍认为，AI带来的制造产能竞争将进一步提升台积电的全球地位和市值表现。

相关链接：

• 台积电市值突破1万亿美元
• 台积电2025年Q2财报创新高

TSMC's Market Cap Breaks $1 Trillion Amid AI Wave, Reaching Historic High 🚀

Briefing:

• TSMC's market capitalization recently surpassed $1 trillion for the first time in the Taiwanese stock market, primarily benefiting from strong Artificial Intelligence (AI) demand and an upgraded performance outlook.
• The company, which manufactures chips for giants like Apple and Nvidia, has seen its stock price surge nearly 50% from its April low, propelled by the expansion of the AI industry. Its Q2 2025 financial report showed record-breaking revenue and net profit. 📈
• TSMC also revised up its full-year 2024 revenue growth forecast to approximately 30%. Analysts from Goldman Sachs and JPMorgan Chase are optimistic about the sustained strength of TSMC's advanced process technology and AI customer spending.
• The market generally believes that the manufacturing capacity competition driven by AI will further enhance TSMC's global position and market performance. 💰

Related Links:

• TSMC's Market Value Surpasses $1 Trillion
• TSMC's Q2 2025 Financial Report Hits New Highs

京东开源多智能体系统JoyAgent-JDGenie，GAIA基准测试精准率创新高

简报：

• 京东正式开源端到端通用多智能体系统JoyAgent-JDGenie，在GAIA基准测试中以75.15%的准确率超越OWL、OpenManus等竞品，成为行业领先。
• JoyAgent-JDGenie支持多模态输入与输出，具备任务分解、智能体协同与记忆优化等功能，适合多场景AI应用开发。
• 框架完全开源（Apache2.0），支持在多平台快速部署，并通过模块化设计便于开发者扩展与定制。

相关链接：

• JoyAgent-JDGenie开源详情
• GAIA基准测试与系统能力分析
• JoyAgent-JDGenie项目GitHub地址

JD Open-Sources Multi-Agent System JoyAgent-JDGenie, Achieves Record-High Accuracy in GAIA Benchmark Test 🚀

Brief:

• JD has officially open-sourced its end-to-end general multi-agent system, JoyAgent-JDGenie. It has set a new industry benchmark by surpassing competitors like OWL and OpenManus with an impressive 75.15% accuracy rate in the GAIA benchmark test. ✨
• JoyAgent-JDGenie supports multi-modal input and output, featuring functionalities such as task decomposition, agent collaboration, and memory optimization, making it ideal for multi-scenario AI application development. 🧠
• The framework is fully open-source (Apache 2.0), supports rapid deployment across multiple platforms, and its modular design facilitates easy extension and customization for developers.

Related Links:

• JoyAgent-JDGenie Open-Source Details
• GAIA Benchmark Test and System Capability Analysis
• JoyAgent-JDGenie Project GitHub Address

英伟达全新轻量推理模型OpenReasoning-Nemotron发布，支持普通游戏电脑高效解题

简报：

• 2025年7月20日，英伟达正式推出OpenReasoning-Nemotron推理模型套件，包含四个基于Qwen-2.5微调的不同参数规模模型（1.5B、7B、14B、32B），源自DeepSeek R1 0528大模型，通过“蒸馏”实现高效压缩。
• 该模型无需强化学习、只用纯监督微调，利用NeMo Skills生成的500万数学、科学及编程数据集进行训练，支持标准游戏电脑运行，业界性能接近主流先进模型，极大降低本地推理门槛和硬件成本。
• 32B模型在AIME24和HMMT 2月数学竞赛获得89.2分和73.8分，即便最小1.5B型号也具较强推理力；所有模型检查点即将在Hugging Face平台开放下载。
• 模型新设“GenSelect模式”，可针对一道题输出多种解答选择最优方案，32B版本多项测试性能已达或超越OpenAI o3-high标准。
• 针对高校及开发者，这套模型支持本地高水平推理实验和定制优化，被英伟达定位为科研探索有力工具，推动AI社区创新研究。

相关链接：

• 英伟达推出 OpenReasoning-Nemotron 推理模型，普通游戏电脑也能玩转高级推理
• 黄仁勋力赞 DeepSeek，称中国创新的步伐不可能被阻挡

NVIDIA Launches New Lightweight Inference Model OpenReasoning-Nemotron, Enabling Efficient Problem Solving on Regular Gaming PCs

Briefing:

• On July 20, 2025, NVIDIA officially launched the OpenReasoning-Nemotron inference model suite, comprising four Qwen-2.5 fine-tuned models of varying parameter sizes (1.5B, 7B, 14B, 32B). Derived from the DeepSeek R1 0528 large model, these models achieve efficient compression through "distillation."
• These models forgo reinforcement learning, relying solely on supervised fine-tuning. Trained on 5 million math, science, and programming datasets generated by NeMo Skills, they can run on standard gaming PCs. Their performance is comparable to mainstream advanced models, significantly lowering the barrier for local inference and reducing hardware costs. 🚀
• The 32B model achieved scores of 89.2 in AIME24 and 73.8 in the HMMT February math competition. Even the smallest 1.5B variant demonstrates strong reasoning capabilities. All model checkpoints will soon be available for download on the Hugging Face platform. 🧠
• A new "GenSelect Mode" has been introduced, allowing the model to output multiple solutions for a given problem and select the optimal one. The 32B version has met or exceeded OpenAI's o3-high standard in multiple performance tests.
• For universities and developers, this model suite supports high-level local inference experiments and custom optimization. NVIDIA positions it as a powerful tool for scientific research and exploration, fostering innovative research within the AI community. ✨

Related Links:

• NVIDIA Launches OpenReasoning-Nemotron Inference Model, Enabling Advanced Reasoning on Regular Gaming PCs
• Jensen Huang Praises DeepSeek, States China's Pace of Innovation Cannot Be Halted