![[0966f859-0132-442c-b8ad-a2c2bf5ca511.mp3]]

MiniMax发布混合架构开源模型M1，长文本处理能力引发关注

简报：

• 上海AI独角兽MiniMax正式推出全球首个开源大规模混合架构推理模型MiniMax-M1，具备百万级Token的长文本处理能力，单次可处理整本英文原著级文本，技术表现领先同类开源模型。
• MiniMax-M1在权威榜单中位居全球开源模型第二名，仅次于DeepSeek-R1-0528，并在工具调用等方面具有明显优势。
• 该模型强化学习阶段研发总投入仅53.5万美元，远低于行业平均，展现出较高研发效率。
• M1发布后，MiniMax还连续发布多款创新AI产品，涵盖视频生成、智能体和音色设计等领域。
• MiniMax自2022年底前就布局AGI，坚定自主基础模型研发，成功实现技术突破。

相关链接：

• MiniMax混合架构模型M1亮相，百万Token长文本处理能力引关注
• MiniMax发布首款推理模型，训练成本远低于业内水平
• 中国MiniMax发布AI推理模型，宣称超越DeepSeek
• MiniMax-M1模型再创新高，开源AI领域取得新突破

MiniMax Releases Hybrid Architecture Open-Source Model M1, Long Text Processing Capability Garners Attention

Briefing:

• Shanghai AI unicorn MiniMax officially launched MiniMax-M1, the world's first open-source large-scale hybrid architecture inference model. It boasts million-token long text processing capabilities, capable of handling entire English novel-level texts in a single pass, outperforming similar open-source models. 🚀
• MiniMax-M1 ranks second globally among open-source models on authoritative lists, trailing only DeepSeek-R1-0528, and demonstrates significant advantages in areas like tool invocation.
• The total R&D investment for the reinforcement learning phase of this model was only $535,000, significantly lower than the industry average, showcasing high R&D efficiency. 💰
• Following the M1 release, MiniMax has successively launched several innovative AI products, covering areas such as video generation, AI agents, and voice design.
• MiniMax has been strategically developing AGI since before the end of 2022, firmly committed to independent fundamental model R&D, and has successfully achieved technological breakthroughs. ✨

Related Links:

• MiniMax Hybrid Architecture Model M1 Debuts, Million-Token Long Text Processing Capability Draws Attention
• MiniMax Releases First Inference Model, Training Cost Far Below Industry Average
• China's MiniMax Releases AI Model That It Claims Beats DeepSeek
• MiniMax-M1 Model Reaches New Heights, Achieving New Breakthroughs in Open-Source AI Field

台积电因美政策调整优先投资美国亚利桑那，推迟日本新厂建设

简报：

• 台积电被曝推迟日本熊本第二座晶圆厂的建设，主要原因是其正加快对美国亚利桑那州凤凰城工厂的资金投入，以应对特朗普政府可能加征关税的风险。
• 美国“大而美”法案预计将半导体企业投资税收抵免从25%提升至35%，且设定了2026年投资优惠截止时限，台积电为抢搭政策红利，加大对美投资，凤凰城基地未来将扩展为多达九座晶圆厂。
• 台积电此举为应对地缘政治与美国最新政策变化，暂缓扩大日本产能，集中资源布局美国市场。

相关链接：

• 台积电传推迟日本新厂建设 投资加码美国以应对关税风险
• 台积电调整全球布局 为美政策踩油门延后日本厂时程

TSMC Prioritizes US Arizona Investment Due to Policy Shifts, Delays New Japan Fab Construction

Brief:

• TSMC is reported to have postponed the construction of its second wafer fab in Kumamoto, Japan. This decision primarily stems from the company accelerating its capital expenditure on the Phoenix, Arizona facility in the U.S. to mitigate the risk of potential tariffs under a possible Trump administration. 💸
• The U.S. "Big and Beautiful" bill is expected to increase investment tax credits for semiconductor companies from 25% to 35%, with a deadline for these incentives set for 2026. To capitalize on these policy benefits, TSMC is significantly boosting its investment in the U.S., with the Phoenix site slated to expand into as many as nine wafer fabs in the future. 🚀
• TSMC's move is a strategic response to geopolitical dynamics and recent U.S. policy changes. By temporarily slowing down its expansion in Japan, the company is consolidating resources to bolster its presence in the American market. 🇺🇸

Related Links:

• TSMC Reportedly Delays New Japan Fab Construction, Boosts US Investment to Counter Tariff Risks
• TSMC Adjusts Global Layout, Accelerates US Plant Timeline Due to US Policy, Delays Japan Plant

Meta洽购PlayAI声音克隆技术，拓展AI语音能力

简报：

• 据消息人士透露，Meta正在与声音克隆初创公司PlayAI进行深入谈判，计划收购其技术及部分员工，以增强AI助手和智能硬件的语音功能。
• 此举是Meta持续加大AI投入、引入高端技术和人才战略的一部分，目的是提升与OpenAI、谷歌等企业在AI领域的竞争力。
• 交易尚未最终敲定，后续可能发生变化，同时Meta仍在积极招聘AI领域相关人员。

相关链接：

• Meta计划收购声音克隆企业PlayAI部分资产，继续为AI业务招聘

Meta in Talks to Acquire PlayAI Voice Cloning Tech, Expanding AI Voice Capabilities

Briefing:

• Sources indicate that Meta is in advanced talks with voice cloning startup PlayAI to acquire its technology and some personnel, aiming to enhance voice functionalities for its AI assistants and smart hardware. 🗣️
• This move is part of Meta's ongoing strategy to increase AI investment, onboard high-end technology and talent, with the goal of boosting its competitiveness against companies like OpenAI and Google in the AI domain. 🚀
• The deal is not yet finalized and could change. Meanwhile, Meta continues to actively recruit personnel in the AI field. 💼

Related Links:

• Meta Plans to Acquire Partial Assets of Voice Cloning Company PlayAI, Continues Recruiting for AI Business

昆仑万维推出第二代开源奖励模型Skywork-Reward-V2，斩获多项评测榜首

简报：

• 2025年7月4日，昆仑万维发布并开源了第二代奖励模型Skywork-Reward-V2系列，共包含8个参数规模从6亿到80亿不等的模型。
• 该系列模型在七大主流奖励模型评测榜单中均获得领先，成为开源奖励模型领域的焦点。
• Skywork-Reward-V2采用包含4000万对偏好对比的混合数据集，并通过人机协同、多轮筛选与自动化处理，实现偏好数据的高质量和大规模平衡。
• 在Reward Bench v1/v2、PPE Preference & Correctness、RMB、RM-Bench、JudgeBench等基准测试中，Skywork-Reward-V2系列全面达到当前最优（SOTA）水平。
• 最大规模模型Skywork-Reward-V2-Llama-3.1-8B在所有主流基准测试中表现最佳，具备优异的泛化能力和多维人类偏好覆盖。
• HuggingFace与GitHub已同步开放模型下载和使用。

相关链接：

• Skywork-Reward-V2模型介绍（TMCnet）
• HuggingFace项目页
• GitHub源码页

Kunlun Wanwei Launches Second-Generation Open-Source Reward Model Skywork-Reward-V2, Securing Top Ranks in Multiple Benchmarks

Briefing:

• On July 4, 2025, Kunlun Wanwei released and open-sourced its second-generation reward model series, Skywork-Reward-V2, comprising 8 models with parameter scales ranging from 600 million to 8 billion.
• This model series has achieved leading positions across seven major reward model evaluation benchmarks, becoming a focal point in the open-source reward model domain. 🌟
• Skywork-Reward-V2 utilizes a hybrid dataset containing 40 million pairs of preference comparisons, achieving high-quality and large-scale balance of preference data through human-machine collaboration, multi-round screening, and automated processing.
• In benchmark tests such as Reward Bench v1/v2, PPE Preference & Correctness, RMB, RM-Bench, and JudgeBench, the Skywork-Reward-V2 series has comprehensively achieved State-of-the-Art (SOTA) performance. 🚀
• The largest model, Skywork-Reward-V2-Llama-3.1-8B, demonstrated the best performance across all mainstream benchmarks, exhibiting excellent generalization capabilities and multi-dimensional human preference coverage.
• Model downloads and usage are now synchronously available on HuggingFace and GitHub. ✨

Related Links:

• Skywork-Reward-V2 Model Introduction (TMCnet)
• HuggingFace Project Page
• GitHub Source Code Page

京东物流推出L4级别自动驾驶无人轻卡，提升物流效率与成本优化

简报：

• 京东物流在第十七届国际交通技术与设备展览会上，正式发布自主研发的无人轻卡“京东物流VAN”，具备L4级别公开道路自动驾驶能力。
• 该车型拥有24立方米超大载货空间和400公里满载续航，成为目前载货量最大的物流无人轻卡，预计可取代传统4.2米货车在摆渡和传站环节的应用。
• 京东物流VAN可自主规划路线，识别并避让障碍，适用于电商仓储、快递分拣、制造业供应链和冷链运输等多种场景，预计能节省约60%成本，并显著提升运营效率。
• 该产品发布标志着京东物流在智能化、自动化物流领域的又一重要突破。

相关链接：

• 京东物流发布自研无人轻卡“京东物流VAN”，L4级别公开道路自动驾驶
• 京东物流无人轻卡VAN亮相，拥有400km续航和24立方米空间

JD Logistics Launches L4 Autonomous Driving Unmanned Light Truck, Enhancing Logistics Efficiency and Cost Optimization

Brief:

• At the 17th International Exhibition on Transport Technology & Equipment, JD Logistics officially unveiled its independently developed unmanned light truck, the 'JD Logistics VAN', featuring L4-level autonomous driving capabilities on public roads. 🚚
• This vehicle boasts an impressive 24 cubic meters of cargo space and a 400-kilometer range when fully loaded, making it the logistics unmanned light truck with the largest cargo capacity currently available. It is expected to replace traditional 4.2-meter trucks for shuttle and transfer operations. 📦
• The JD Logistics VAN can autonomously plan routes, identify and avoid obstacles, making it suitable for various scenarios including e-commerce warehousing, express sorting, manufacturing supply chains, and cold chain transportation. It is projected to save approximately 60% of costs and significantly enhance operational efficiency. 💰
• The launch of this product marks another significant breakthrough for JD Logistics in the intelligent and automated logistics sector.

Related Links:

• JD Logistics Unveils Self-Developed Unmanned Light Truck 'JD Logistics VAN', Featuring L4 Autonomous Driving on Public Roads
• JD Logistics Unmanned Light Truck VAN Debuts, Boasting 400km Range and 24 Cubic Meters of Space