早报｜iPhone Duo限购两台/雷军：汽车没有安全其他都是0/宝马员工争相离职挤崩系统
📱iPhone Duo 开启全球限购，每位顾客最多买 2 台
🛡️PHP 紧急修复两个严重漏洞，攻击者无需登录即可远程执行代码
🏭三星、SK 海力士升级中国工厂，拟转产更高层数 NAND
🚗雷军：汽车产业如果没有安全，其他都是 0
🏢宝马员工争相离职挤崩系统，个别方案补偿超 65 万欧元
💵Today AI 累计融资超 2 亿美元，腾讯、蚂蚁等参投
🧠OpenAI 模型将入驻谷歌云，企业可用云消费承诺购买
💾英伟达把 AI 工厂装进集装箱，DSX POD 最快 14 天部署
💼中办、国办：全面实施「人工智能+」行动，支持企业主导成果转化
💬曝DeepSeek 近 1000 亿元新融资接近完成
✨Codex 推出永久对话缓存
📉特斯拉三季度净利润下降 43%，自由现金流转正
🏢IBM 软件业务增长回升，上调全年自由现金流预期
💵代码越写越快，测试需求激增：Blacksmith 估值达 10 亿美元
💡何小鹏：自动驾驶应跳过 L3，安全性还要打问号
🧠Anthropic 推出 Claude Duet，会主动汇报并在后台继续办事
📱国补 3149 元起，荣耀 WIN2 系列开售
🪄Flowtica 获新融资，磁吸墨水屏硬件 Verso 计划 12 月出货
✨戴森推出水光直板夹，拉直头发时用冷凝管降温
📹生数科技公开 Motubrain 2.0 训练流程，计划连续直播半个月
🔋宇树发布 H2-X「功夫传奇」，最长续航约 8 小时
📷徕卡 Q3 43 推出 F1 特别版，碳纤维握柄搭配专属赛车色彩
🪄《绝地求生》厂商开放 Raon-1 mini 模型，量化版可在单张 RTX 5090 运行
💾DGX Spark 可组成 8 台集群，统一内存最高达 1TB
☕星巴克 10 月 22 日起全国实行「消费入座」，将投入超 3 亿元升级服务
🧠美国沃尔玛为 AI 代购设独立档案，用户可限额并随时撤销授权
🎬《沙丘 3》内地定档 12 月 31 日
👀《欢乐合唱团》将重启，Disney+ 与 Hulu 正式预订新剧
🎨汤姆·克鲁斯将获 AFI 终身成就奖，明年 4 月举行致敬典礼
iPhone Duo 开启全球限购，每位顾客最多买 2 台
苹果官网显示，原定 10 月 23 日发售的 iPhone Duo 目前每位顾客限购 2 台。
iPhone Duo 于 10 月 16 日晚 8 点在苹果官网、Apple Store App、苹果天猫旗舰店等线上渠道开启预售。预售期间涌入量过大，平台一度卡顿。报道称，首发当日到店取货库存已售罄，部分线上订单的配送时间已推迟至明年 1 月下旬。该机国行起售价为 15999 元。
二级市场上，据中新经纬报道，预售开启几分钟后，得物平台上 iPhone Duo 多个版本的成交价突破 2 万元。其中 1TB 黑色版本成交均价为 28642 元，比官方售价高出 7143 元；256GB 黑色版本成交均价为 22085 元，上涨 6086 元，涨幅约 38.0%。
不过，目前第三方渠道报价已大幅回落，多位黄牛在二手交易平台上仅加价 3000 元，并承诺首销当天拿货。
PHP 紧急修复两个严重漏洞，攻击者无需登录即可远程执行代码
PHP 开发团队发布紧急安全更新，修复 CVE-2026-109618 和 CVE-2026-109619 两个严重漏洞。攻击者可能通过特制网络请求，在无需身份验证的情况下远程执行代码；官方确认，在 Apache mod_php 场景中，攻击甚至可在应用代码运行前触发。
攻击者可能无需账号或密码，只需向受影响的服务器发送特制请求，就能远程执行代码。漏洞存在于 PHP 核心请求处理流程中，不依赖 WordPress、Drupal 等网站应用本身存在缺陷。
一旦成功利用漏洞，就可能在 PHP 工作进程的权限范围内执行代码、修改网站，或读取进程内存中的密码、身份验证令牌和 API 密钥等敏感信息。
修复版本为 PHP 8.5.12、8.4.27、8.3.36 和 8.2.35。官方强调，临时缓解措施只能降低部分攻击路径的风险，更新才是完整解决办法；PHP 8.1 及更旧版本已不再获得项目常规安全更新，使用者需核对发行版或长期支持供应商的修复包。
OpenAI 表示，Calif 在 Patch the Planet 计划中借助 GPT-5.6 Sol 发现其中一处漏洞，Wiz 也独立发现并报告了两处问题。维护者与研究团队、基础设施及安全服务商已协调约两个月，为披露前的修复和防护做准备。
三星、SK 海力士升级中国工厂，拟转产更高层数 NAND
据路透社报道，三星和 SK 海力士正升级各自在中国的 NAND 闪存工厂，以应对长江存储的竞争及 AI 数据中心需求。两家公司均确认正在安装或升级设备，并称与美国政府保持沟通。
三星西安工厂拟转产 286 层 V9 NAND。知情人士称，该项目计划今年 12 月实现每月 5000 片晶圆产出，2027 年 12 月增至 4 万片。西安工厂约占三星全球 NAND 产量的 40%，目前主要生产约 200 层产品。三星强调，这次属于技术升级，不涉及新增产能或建设新工厂。
SK 海力士则拟将大连工厂的产品从 192 层升级至 240 层或以上。一名知情人士称，公司可能进一步申请在当地研发 300 层以上技术，为未来在美国生产做准备。
两家公司在中国安装美国设备需逐年申请许可。路透尚未确认美国已批准此次升级所需设备的具体范围；美国商务部重申，此前政策支持现有工厂运行，但不支持扩产或技术升级。
雷军：汽车产业如果没有安全，其他都是 0
10 月 21 日，小米汽车北京设计中心正式启用。雷军同日在世界智能网联汽车大会上表示，汽车产业如果没有安全，其他都是 0；所有的智能性能和创新设计都失去了意义。小米从造车第一天起就高度重视安全。
谈到智能网联汽车的下一阶段，雷军认为，竞争将更多围绕生态展开。他提出，中国拥有完备的产业链、活跃的 AI 创新氛围和庞大的单一市场，有机会把整个智能生态连接起来，走出中国式的创新路径。
雷军以小米的消费级 IoT 平台为例称，截至二季度末，设备连接数已超过 11 亿台。为了完善生态体验，小米持续投入芯片、操作系统和 AI 具身智能等底层技术；过去 5 年硬核科技研发投入为 1055 亿元，预计未来 5 年相关投入将超过 2000 亿元。
小米澎程官方账号称，澎程汽车设计展已面向公众开幕，北京汽车设计中心也迎来建成后的首次对外开放。雷军随后转发博文，确认设计中心已正式启用，并表示计划前往参观。
宝马员工争相离职挤崩系统，个别方案补偿超 65 万欧元
宝马的自愿离职补偿被「挤爆」，已提前停止接受新申请。据德国《商报》10 月 20 日报道，自 10 月 8 日开放以来，已有数千名员工表达离职意愿，宝马确认目标已提前达到，计划在下月中旬前完成方案处理。
此前，数万名员工收到补偿邀请，大量员工同时登录查看个人方案，曾使内部人力资源门户瘫痪数小时，系统于次日恢复。员工表达意愿不等于获准离职，这项计划要求员工与雇主双方自愿同意。
《商报》披露的个别补偿方案超过 65 万欧元，主要涉及资历较深、薪资较高的员工，不能视为普遍标准。Motor1 的完整报道显示，方案可结合一次性付款、分期、养老金及保险等方式，部分员工离职后仍可保留购车优惠。
宝马目标是在 2027 年底前减少约 8000 个岗位。
Today AI 累计融资超 2 亿美元，腾讯、蚂蚁等参投
据雷递报道，Today AI 于 10 月 21 日宣布，团队累计完成超过 2 亿美元融资，投资方包括 IDG 资本、腾讯、五源资本、蚂蚁集团、红杉中国、传音控股、同歌创投及上海交大菡源资产。
Today AI 由 Teambition 创始人、前豆包 PC 端负责人齐俊元创办，主打个性化、主动式 AI 助理。产品已在国内上线，可承担日程管理、资料研究、工作协助和随身记录等任务。
据智东西报道，Today 支持 iOS、安卓、Windows、Mac 和 Linux，并可连接企业微信、微信读书、百度网盘、金山文档、腾讯会议等应用。团队希望通过持续交互积累用户记忆，在用户提出需求前提供协助。
OpenAI 模型将入驻谷歌云，企业可用云消费承诺购买
OpenAI 与 Google Cloud 宣布合作，将前沿模型和 Codex 智能体工作流接入谷歌云的 Model Garden，计划在未来几周全面开放。
企业可通过 Google Cloud Marketplace 购买这些模型，费用可抵扣既有的谷歌云消费承诺。双方将把模型能力与谷歌云的访问控制、集中式数据治理和基础设施结合，用于软件开发及复杂的智能体任务。
OpenAI 还表示，其模型将很快进入 Google Antigravity 和 Gemini Agent，让企业在这些工具中选择不同公司的模型。微软仍是 OpenAI 的主要云合作伙伴。
英伟达把 AI 工厂装进集装箱，DSX POD 最快 14 天部署
英伟达在 GTC 柏林大会推出 DSX POD：由合作伙伴在模块化集装箱内建造完整的 Vera Rubin AI 工厂，出厂前完成集成和验证，计划 2027 年初在全球交付。
按英伟达的说法，在场地电力、冷却、网络等配套已准备好的前提下，这套方案可将现场部署时间从 6 个月缩短至最快 14 天。计算、网络、布线、机架供电和液冷的安装及测试被前移到制造工厂，与场地准备并行进行。
参与伙伴包括戴尔、惠普企业、联想、施耐德电气和维谛技术等。黄仁勋表示，AI 工厂开始产生价值的节点是投入模型训练或生产推理、生成 token，DSX POD 希望压缩的正是交付后到正式运行之间的时间。
中办、国办：全面实施「人工智能+」行动，支持企业主导成果转化
中共中央办公厅、国务院办公厅发布《关于推动科技创新和产业创新深度融合的意见》，提出全面实施「人工智能+」行动，推动人工智能全方位赋能千行百业。
《意见》要求加速工业互联网创新发展，以先进技术改造传统产业；在量子科技、脑机接口、具身智能、6G 等前沿领域探索技术路线、应用场景和商业模式，推进产业化、商业化。
在成果转化上，产业应用目标明确的国家科技计划项目原则上由企业牵头实施并主导成果转化，鼓励高校、科研院所按照先使用后付费方式向中小微企业许可科技成果。同时，文件提出推动健全人工智能、生物技术等前沿领域法律法规，探索「沙盒监管」方式。
曝DeepSeek 近 1000 亿元新融资接近完成
据雷递报道，DeepSeek 新一轮近 1000 亿元融资接近完成，投前估值约 5000 亿元。报道援引知情人士称，砺思星深、拾象尚势两只基金均参与本轮投资，腾讯也继续跟投。
完美世界披露的基金进展公告显示，砺思星深已募集完毕，总规模为 75.2 亿元，公司认缴 5000 万元；拾象尚势基金总规模为 36.5 亿元，公司认缴 1 亿元。
按雷递获得的信息，本轮新进投资者以产业资本为主，包括石溪资本、中芯聚源、CPE 源峰、君联资本和吉利等，宁德时代、腾讯、IDG 资本、砺思资本和拾象资本等老股东继续跟投。
Codex 推出永久对话缓存
OpenAI 产品负责人 Tibo Sottiaux 宣布，API 现可选择 12 小时提示缓存；对 Codex 和 Work 的已登录用户，公司会永久保留缓存。用户回到旧对话，或中断数小时后继续，也可命中缓存。
OpenAI Developers 的公告解释，长期运行的智能体会在一天中反复使用相同上下文。新选项可让共享指令、工具和上下文保留更久，即使两次请求间隔数小时，仍可按较低的缓存输入费率计费。
Tibo 将这项变化列为「第 16 天」更新，并表示恢复旧会话不会受到额外惩罚。
特斯拉三季度净利润下降 43%，自由现金流转正
特斯拉公布 2026 年第三季度财报，营收 280.46 亿美元，与上年同期基本持平；归属普通股股东的 GAAP 净利润为 7.87 亿美元，同比下降 43%。公司当季交付 486532 辆车，同比下降 2%。
几组数据呈现了本季的业务与支出变化：
- 汽车业务营收 205.60 亿美元，同比下降 3%；储能业务营收 28.81 亿美元，下降 16%；服务及其他业务营收 46.05 亿美元，增长 33%。
- GAAP 营业利润为 7.18 亿美元，同比下降 56%，营业利润率为 2.6%；摊薄每股收益为 0.22 美元，非 GAAP 摊薄每股收益为 0.48 美元。
- 资本开支为 47.32 亿美元，低于上季度的 57.89 亿美元；经营现金流为 62.63 亿美元，自由现金流为 15.31 亿美元，由上季度负值转正。
公司称，北美新车交付中超过四分之三附带 FSD 订阅，全球活跃 FSD 订阅数达 174 万，口径包括一次性购买与月订阅，不含免费试用。特斯拉同时强调，FSD 仍需驾驶员主动监督。Cybercab 已于 9 月 3 日在奥斯汀加入 Robotaxi 服务，Tesla Semi 和 Optimus 也已分别在内华达和弗里蒙特开始生产。
IBM 软件业务增长回升，上调全年自由现金流预期
IBM 公布 2026 年第三季度业绩，营收 171.03 亿美元，同比增长约 5%，按固定汇率计算增长 6%。持续经营业务 GAAP 净利润为 21.99 亿美元，增长 26%；相应摊薄每股收益为 2.31 美元，经营口径非 GAAP 每股收益为 2.95 美元。
软件业务营收约 78 亿美元，同比增长 8%。其中，Red Hat 增长 11%，数据业务增长 19%，自动化业务下降 3%。咨询业务营收约 53 亿美元，与上年同期基本持平；基础设施业务营收约 37 亿美元，增长 4%，但其中 IBM Z 大型机收入下降 20%。
公司当季经营现金流约 40 亿美元，自由现金流约 30 亿美元，向股东派息约 16 亿美元。前三季度自由现金流累计约 78 亿美元。
IBM 维持全年固定汇率营收增长 4% 至 5% 的预期，将全年自由现金流展望上调为同比增加超过 10 亿美元。董事会另批准每股 1.69 美元的季度现金股息，登记日为 11 月 10 日，支付日为 12 月 10 日。
代码越写越快，测试需求激增：Blacksmith 估值达 10 亿美元
据《财富》报道，代码测试基础设施公司 Blacksmith 完成红杉资本领投的 6600 万美元 C 轮融资，估值达到 10 亿美元。公司成立于 2024 年，为代码合并和部署前的自动构建、测试提供算力与软件。
Blacksmith 称，使用其平台的企业已超过 1 万家，客户包括 Granola、Handshake、Supabase 和 LangChain。每周验证任务从年初的 500 万次增加到目前的 9000 万次，年经常性收入同期增长 17 倍。
联合创始人兼 CEO Aditya Jayaprakash 表示，AI 工具让写代码变得容易，瓶颈随之转向验证代码。Blacksmith 购买自有服务器，在数据中心部署，并通过软件复用测试所需的重复准备工作。
本轮资金主要将用于购买更多算力。公司也面临计算和内存供应紧张，以及模型厂商自行提供代码验证服务的竞争。
💡 何小鹏：自动驾驶应跳过 L3，安全性还要打问号
小鹏汽车董事长何小鹏在世界智能网联汽车大会上表示，自动驾驶应该跳过 L3。他认为，L3 模式将安全责任转移给系统，但用户体验没有得到优化，安全性是否提高也需要打问号。
在他的判断中，L3 本质上更接近 L2 的增强形态，他看好的发展路线是 L2、L4，以及重新定义后的 L5。各国监管体系要支撑自动驾驶向更高等级演进，前提应当是安全、可管控。
何小鹏还提出，过去的 L5 定义主要从技术维度出发，没有充分从安全、场景可控的角度考量。
汽车是在一定条件下提供服务的被动式自动驾驶，机器人则可能需要 24 小时主动执行任务，因此未来 L5 的定义也应随应用场景调整。
谈及落地时间，他区分了技术和政策两条进度：从技术层面，他预计 L4 将在 2028 年下半年落地；从法规和政策角度，他判断时间可能在 2028 至 2032 年之间。
Anthropic 推出 Claude Duet，会主动汇报并在后台继续办事
Anthropic 发布 Claude Duet，将 Claude 做成会主动联系用户、持续处理任务的个人智能体。目前产品开放研究预览候补名单，先面向美国 Claude Pro 和 Max 订阅用户，提供独立 iOS、Android App 及 claude.ai 网页入口。
Duet 基于 Claude Opus 5.5，沿用用户已有的记忆、技能和应用连接。它每天早上汇总已连接应用中的重要事项，白天继续关注变化，只在需要处理或决策时发消息；用户也可交给它安排会议、制作预测表、搭建仪表盘或开发网页应用等任务，离开后它仍会继续工作。
它在云端电脑里使用用户连接的工具，产品页称支持超过 6000 个连接器和插件。涉及发送消息、付款等敏感操作时需先取得确认；Anthropic 表示，Duet 不直接看到密码和登录令牌，还加入专门的信息保护分类器，限制身份证件、信用卡等敏感信息外传。
研究预览包含在 Pro 和 Max 套餐内，与其他 Claude 使用共享额度，后台检查也会消耗额度；Team 和 Enterprise 暂不包含。官方提示，产品仍可能误读对话、判断错紧急程度，或漏掉两次检查之间到来的消息。
国补 3149 元起，荣耀 WIN2 系列开售
荣耀 10 月 21 日在杭州发布 WIN2、WIN RT2 两款手机，并于发布当天开售，首销国补到手价分别为 4999 元起、3149 元起。本场还带来 WIN Turbo2、WIN 平板及 WIN 耳机。
- 荣耀 WIN2：搭载第六代骁龙 8 超级至尊版，配备双风扇立体风网散热系统，总风量为 1.19CFM；内置 10100mAh 电池，支持 100W 有线和 80W 无线充电。另有 2 亿像素主摄、USB 3.2、3D 超声波指纹和可自定义的电竞 WIN 键。
- 荣耀 WIN RT2：搭载第五代骁龙 8 至尊版，配备与 WIN2 同款的天王星风扇，电池增至 12070mAh。两款手机均采用 2K 185Hz 双芯绿洲屏，以及 LPDDR5X Ultra 内存和 UFS 4.1 闪存。
- 荣耀 WIN Turbo2：已开启预售，将于 10 月 30 日 10:08 开售，国补到手价 2251.65 元起。
- 荣耀 WIN 平板：采用 8.4 英寸 240Hz OLED 屏，搭载第五代骁龙 8 至尊版、风冷散热系统和 10000mAh 电池。
- 荣耀 WIN 耳机：支持 40ms 低延迟和 55dB 降噪，并通过 Hi-Res 认证。
WIN2 还将推出《三角洲行动》职业赛事联名款，10 月 30 日开启电商预约，计划 11 月登场。
Flowtica 获新融资，磁吸墨水屏硬件 Verso 计划 12 月出货
据硬氪报道，AI 硬件公司 Flowtica 已完成新一轮数千万元融资，投资方包括祥峰资本、源码资本、五源资本和清水湾基金。其磁吸式墨水屏硬件 Verso 本周开始预收定金，首批量产产品计划于 12 月出货。
Verso 可贴在手机背面，厚 6.5mm、重 78g，配备常显墨水屏、实体按键、Type-C 接口、四麦克风阵列及骨传导传感器。屏幕用于呈现日程、待办及 AI 提炼的信息，侧边按键支持盲操录音和调用 AI。
创始人 Rian 在采访中解释，语音适合输入指令，屏幕便于快速扫读结果。按团队的产品设计，会议记录将与个人长期记忆联动，再生成待办拆解、会前简报和跟进草稿；公司此前已推出录音笔 Scribe，Verso 是其第二款硬件。
戴森推出水光直板夹，拉直头发时用冷凝管降温
戴森在中国大陆推出 Corrale CoolShine 水光直板夹，在柔性弹板之外加入可拆卸的散热冷凝管，让头发经过热夹板塑型后同步降温。
据戴森提供的产品材料，配件采用密封铜管，内部少量水在蒸发过程中带走热量，减少热量停留在发丝上的时间。柔性弹板则通过形变包裹并聚拢发束，让头发均匀受力，减少反复拉夹。
新机提供 145℃、165℃、185℃ 和 210℃ 四档温度，约 15 秒完成预热，含配件净重约 390 克，并支持全球通用电压。戴森称，与未经造型的头发相比，造型后光泽度最高可提升 154%。
产品于 10 月 20 日在中国大陆发售，提供晨曦粉和晶耀紫两种配色。
生数科技公开 Motubrain 2.0 训练流程，计划连续直播半个月
生数科技世界模型团队宣布，10 月 21 日开启 Motubrain 2.0 训练全流程公开直播，计划持续半个月，展示具身模型从训练、测试到能力迭代的过程。团队表示，Motubrain 2.0 即将开源，目前尚未公布具体开源日期。
Motubrain 系列面向机器人任务，研究不同任务、场景和机器人本体之间的能力泛化。这次直播覆盖四个环节：预训练阶段观察训练规模对性能与零样本泛化能力的影响；后训练阶段提高具体任务的执行能力；Agent 任务执行阶段探索已有经验向其他任务迁移；递归自我改进（RSI）阶段则研究模型能否依据执行反馈，评估并改进自身行动。
团队称，相关研究仍处于探索阶段，具体效果有待后续实验与评测验证。官方 Live Lab 页面已展示训练损失、训练步数、模型评测曲线及 4 路摄像头入口，评测任务包括挂耳机、整理抽屉、拾取交接物品和擦白板。
宇树发布 H2-X「功夫传奇」，最长续航约 8 小时
昨天，宇树科技发布人形机器人 H2-X「功夫传奇 守护」。
据介绍，H2-X 以强劲性能提供服务和守护，最长续航时间约 8 小时，并提醒使用者安全、友好地使用机器人。该视频封面中的型号为 Unitree H2-X。
宇树官方此前公布，2026 世界物理 AI 大会暨宇树合作伙伴生态大会将于 10 月 22 日举行，H2-X 在大会前一天公开。
徕卡 Q3 43 推出 F1 特别版，碳纤维握柄搭配专属赛车色彩
徕卡与 Formula 1 宣布授权合作，推出 Q3 43 Formula 1 特别版相机，10 月 21 日起通过徕卡门店、官网商店及专业零售商发售。
特别版将赛车材质带到相机上：机身包覆与赛车方向盘类似的黑色防滑 Microvelour 材料，握柄、指托和软快门按钮采用德国手工制作的碳纤维部件。顶盖、镜头盖及背带带有 F1 与徕卡标识，软快门按钮和指托右侧功能键则使用 F1 红色。
相机还新增专属「Formula 1」Leica Look，通过更高对比度与饱和度呈现赛道氛围。用户可经 Leica FOTOS App 将色彩风格传到相机，拍摄时就在取景器中预览效果，并直接应用于 JPEG。
核心规格延续标准版：6000 万像素全画幅 CMOS、Maestro IV 处理器、APO-Summicron 43mm f/2 ASPH. 镜头及最高 8K 视频录制。配套碳纤维握柄支持无线充电和 Arca-Swiss 快装结构。
《绝地求生》厂商开放 Raon-1 mini 模型，量化版可在单张 RTX 5090 运行
《绝地求生》开发商 KRAFTON 公开自研语言模型 Raon-1 mini 预览版，重点面向韩语对话、编程和智能体任务。模型总参数 288 亿，每次处理 token 激活约 43 亿参数，原生支持 262144 token 上下文。
它采用混合注意力架构：25 层中有 18 层使用基于 Kimi Delta Attention 改进的 RDA，另 7 层为完整注意力；团队还用 Raon-mHC 简化 DeepSeek 的超连接机制，以降低训练计算量。
官方自测中，SWE-Verified 代码修复得分为 70.8，但 Terminal-Bench Lite 为 47.0，不同任务表现存在差异。模型提供 NVFP4 量化版，可在拥有 32GB 显存的单张 RTX 5090 上推理，也支持 llama.cpp 和 Apple 芯片上的 MLX。
下载的模型权重及组件采用 Apache 2.0 许可，需先申请访问；官方托管 API 和体验服务另有条款。这仍是预览版本，规格与测试结果可能调整。
DGX Spark 可组成 8 台集群，统一内存最高达 1TB
英伟达在 GTC 柏林大会更新本地 AI 产品与软件：DGX Spark 现在可手动配置最多 8 台设备组成集群，使用内置 ConnectX-7 网络互联，统一内存总量最高达到 1TB。
这套集群面向小型开发团队，可让多人同时连接同一模型，运行更大模型、更长上下文和多个智能体。NVIDIA Sync Model Launcher 计划本月晚些时候推出，可按单机或集群配置下载适合的模型和容器，并接入 OpenCode。
硬件方面，64GB 版 DGX Spark 将于 10 月 23 日从合作伙伴渠道陆续推出。更高定位的 DGX Station 已提供 748GB 一致性内存，Windows 版计划 12 月由合作伙伴推出，用于在本地模型与制造业专业软件之间建立工作流。
同场上线的十月软件更新还为 DGX Spark 加入 Wi-Fi Assistant，便于设备在无显示器的 appliance 模式下切换网络。
星巴克 10 月 22 日起全国实行「消费入座」，将投入超 3 亿元升级服务
星巴克中国将于 10 月 22 日起在全国实行「消费入座」，门店会摆放相关提示牌。公司表示，会向门店伙伴提供统一服务指引，执行时遵循友好沟通原则，并妥善处理合理的突发需求及特殊情况；遇到孕妇、行动不便者等需要帮助的顾客，会优先给予帮助。
与这项门店空间规则同时公布的，还有新财年服务升级计划。星巴克中国宣布，2027 财年将新增投入超 3 亿元，用于增加高峰时段及臻选吧台人力、升级数智化系统和核心设备，并改造约 1000 家门店。
其中，全国近半数门店将增加高峰时段人力，每月新增超过 20 万小时服务时间，臻选吧台另增超过 7 万小时。新增 70 多项功能的智能移动 POS 机将自 2027 年 1 月起分阶段上线。
服务保障还包括：顾客对饮品不满意可免费重做；下错取单门店时，可凭有效订单由所在门店协助重新制作。
美国沃尔玛为 AI 代购设独立档案，用户可限额并随时撤销授权
沃尔玛宣布为 AI 代购推出 Agent Profiles、Agent Gateway 和专门的智能体使用条款，覆盖沃尔玛与山姆会员商店。用户可以选择自己习惯的 AI 应用代买商品，同时单独管理每个智能体的权限。
Agent Profiles 计划在接下来一个月上线。每个智能体将在顾客账号内拥有独立档案，代购订单与本人订单分开显示；用户可要求逐笔确认，也可设定消费限额，并随时关闭授权。
已进入测试的 Agent Gateway 通过 MCP 协议连接智能体与商店系统，让它们直接查找商品和完成购买，减少模拟浏览器操作造成的卡顿和错误。沃尔玛称，系统可帮助智能体找到用户常买的商品，而不必向它披露完整购物历史。
条款禁止智能体超出顾客意图采取行动、修改顾客支付方式或读取任务所不需要的信息，违规智能体将被封禁。
沃尔玛还表示，同一商品面向 AI 智能体与直接购物顾客展示相同价格，适用优惠及费用会在付款前列明。
《沙丘 3》内地定档 12 月 31 日
《沙丘 3》宣布将于 12 月 31 日在中国内地上映。IMAX 官方同步发布定档信息，称影片采用 IMAX「特制」拍摄，包含 71 分钟独家画幅，并配以定制混音。
影片由丹尼斯·维伦纽瓦执导，是其《沙丘》三部曲的最终章，继续围绕保罗·厄崔迪掌握权力后的命运展开。提莫西·查拉梅与赞达亚饰演的保罗、契妮在故事中走向命运的两端，逝去故人的归来也给保罗带来新的谜题。
影片由传奇影业与华纳兄弟探索集团打造，北美定于 12 月 18 日上映。
《欢乐合唱团》将重启，Disney+ 与 Hulu 正式预订新剧
《欢乐合唱团》（Glee）将以新剧回归。Disney+ 与 Hulu 已正式预订重启版，制作方仍是原剧的 20th Television。
据《好莱坞报道者》报道，主创瑞恩·墨菲在 8 月表示考虑重启，并在 9 月透露已与原剧共同创作者布拉德·法尔查克、伊恩·布伦南写好新版本剧本。此次预订公告尚未公布演员和制作团队，相关主创协议仍在敲定。
原剧于 2009 至 2015 年在 Fox 播出，共 6 季，围绕俄亥俄州一所高中的合唱团展开。它把流行歌曲翻唱、校园成长与人物关系结合，曾获得 6 座艾美奖，原版目前可在 Hulu 观看。
汤姆·克鲁斯将获 AFI 终
本文由 AI图鉴自动抓取并翻译，版权归原作者所有；如需查看原文请点下方链接。
The morning paper iPhone Duo limited the purchase of two/Rattling Army units: the rest of the vehicles were unsafe.
📱iPhone Duo opens the global limit for up to 2 purchases per customer
PHP Emergency repair of two serious faults and remote implementation of code by the assailant without login
Samsung, SK, Hercules upgrade to Chinese plant, proposed for transfer to higher level NAND
Mine: If the automobile industry is not safe, the rest is 0
BMW employees competed for separation and scrambling, and individual programs were reimbursed over €650,000.
Tencent, ants, etc.
🧠Open AI model will be added to Google Clouds, which enterprises can buy using cloud consumption commitments
Yin Weida put the AI plant in a container, DSX POD, 14 days fast.
• Central, National and National Offices: full implementation of the AI+ initiative to support corporate ownership of results
We're in contact with DeepSeek.
ZCodex Launching Permanent Dialogue cache
Net profits declined by 43% in the third quarter, free cash flows went positive.
🏢IBM software business growth rebounded, increasing year-round free cash flow expectations
The faster the code is written, the greater the demand for testing: Blacksmith is valued at $1 billion.
Peng: Autopilot should skip L3. Safety is a question mark.
Anthropic, launch Claude Duet, will report and continue in the background.
Zhong Zhong, 3149 yuan, honors Win2 series.
🪄Flowtica got new financing, Magnetic Ink Screen Hardware, Verso Plan. Goods
Dyson rolls out a water-lighted straightboard and cools his hair with a condenser.
Motubrain 2.0 training process planned to be broadcast live for half a month
publish H2-X "Kung Fu legend" with a maximum duration of 8 hours
Q343 Launch F1 special, carbon-fibre-shareding with exclusive car colour
Open Raon-1 Mini model, quantization version running on a single RTX 5090
DGX Spark can form 8 clusters with a unified memory of up to 1 TB
In October 22nd, the entire country will be running a "consumer seat" and will be investing over $300 million in upgrading services.
The U.S. Wal-Mart has a separate file for AI, which can be limited to users and which can be revoked at any time.
Dec. 31, inside dunes 3
The Happy Choir will be restarted, Disney+ and Hulu officially booking a new play.
Tom Cruz will be awarded the AFI Lifetime Achievement Award for next April.
iPhone Duo opens global limited purchases, up to 2 per customer
Apple Network shows that the iPhone Duo originally scheduled for sale on October 23 is currently limited to two purchases per customer.
The iPhone Duo opened a pre-sale at 8 p.m. on October 16th at AppleNet, Apple Store App, Apple Skycat Flags. The influx was excessive during the pre-sale period, and the platform was once at Carden. It was reported that stocks had been sold out on the first day of distribution and that distribution of orders on some lines had been delayed until late next January. The start-up price of the aircraft is US$ 15999.
On the secondary market, it was reported that a few minutes after the opening of the pre-sale, several versions of the iPhone Duo version of the acquisition platform had been sold at a price of $20,000. Of these, the average price of the 1TB black version was $28642, which was $7143 higher than the official sale price, and the average price of the 256GB black version was $22085, an increase of $6,086 or 38.0 per cent.
At present, however, the third-party-channel offer has fallen significantly, with several cows increasing their prices on second-hand trading platforms by only $3000 and pledging to take the goods on the day of the first sale.
PHP Emergency repair of two serious loopholes and remote implementation of the code by the assailant without login
PHP development team publish emergency security update to repair two serious loopholes, CVE-2026-109618 and CVE-2026-109619. The attacker may request the remote implementation of the code without the need for identification through a specially designed network; the official confirmation is that the attack can be triggered even before the application of the code can run in the Apache mod php scene.
The assailant may not need an account number or password, but simply send a specially designed request to the affected server to perform the code remotely. Gaps exist in the PHP core request processing process and there are deficiencies in web applications such as WordPress and Drupal.
Once a loophole has been successfully exploited, it may be possible to implement codes, modify websites, or read sensitive information such as passwords, authentication badges and API keys stored in the PHP process permission.
The restored versions are PHP 8.5.12, 8.4.27, 8.3.36 and 8.2.35. Officially, it was stressed that interim mitigation measures would only reduce the risk of partial attacks and that upgrading would be the complete solution; PHP 8.1 and older versions were no longer subject to routine project security updates, and users needed to check the release or long-term support for vendor repair packages.
Open AI states that Calif discovered one of these loopholes through GPT-5.6 Sol in the Patch the Planet project, and Wiz independently identified and reported two problems. Maintainers have coordinated with research teams, infrastructure and security providers for approximately two months to prepare for repairs and protection before disclosure.
Samsung, SK Hercules upgraded the Chinese plant to a higher level, NAND.
According to Reuters, Samsung and SK Hercules are upgrading their NAND flash plants in China to meet the competition of Yangtze storage and the needs of AI data centres. Both companies confirmed that equipment was being installed or upgraded and stated that they were in communication with the United States Government.
The Samsung Xian plant is to be converted to the 286th floor V9 NAND. According to the source, the project is planned to produce 5,000 slices per month in December of this year, increasing to 40,000 in December 2027. The Xian plant, which accounts for approximately 40 per cent of the global production of NAND, is currently producing about 200 layers of products. Samsung stressed that this was a technical upgrade that did not involve new capacity or new plants.
For its part, SK Hercules proposes to upgrade the products of the Dalian plant from 192 to 240 layers or more. According to an informed source, the company may further apply for local development of over 300 layers of technology in preparation for future production in the United States.
Both companies are required to apply for annual permits to install United States equipment in China. Reuters had yet to confirm that the United States had approved the specific scope of equipment required for the upgrade; the United States Department of Commerce reiterated that previous policies supported the operation of existing plants but did not support expansion or technological upgrading.
RPA: If the auto industry is not safe, the rest is 0.
On 21 October, the Beijing Design Centre for Mi Motors was officially inaugurated. On the same day, at the World Smart Net Consortium Congress, the Rebels said that, if the automobile industry was not safe, the rest would be zero; all intelligent performance and innovative design would be meaningless. Mi has attached great importance to safety since the first day of car construction.
Turning to the next phase of the Smart Nets, the Army believes that competition will be more ecological. He suggested that China had a full industrial chain, an active AI innovation climate and a large single market, and had the opportunity to connect the entire intelligent ecology and move beyond the Chinese innovation path.
By the end of the second quarter, the number of equipment connections had exceeded 1.1 billion, citing the example of the Iot platform for millet consumption. In order to improve the ecological experience, millet has been continuously investing in ground-level technologies such as chips, operating systems and AI with intelligence; in the last five years, R&D in hard-core technology has been $105,550 billion, and it is expected that over the next five years more than $200 billion will be invested.
According to the official account of Michang, the car design exhibition has been opened for the public and the Beijing Motor Design Centre is open for the first time since its construction. The Army subsequently forwarded Boven, confirming that the design centre had been officially inaugurated and indicating that it planned to visit.
BMW employees competed for separation and scrambling, and individual programs were reimbursed over €650,000
BMW’s voluntary separation compensation was “squeezed” and the new application was withdrawn early. According to the German Journal of Business on 20 October, since its opening on 8 October, thousands of employees have expressed their will to leave, BMW has confirmed that the target has been met ahead of schedule and that the programme is scheduled to be completed by mid-minster next month.
Earlier, tens of thousands of staff had been invited for compensation, and a large number of staff had been able to log in and view individual programmes, paralysing the internal human resources portal for several hours and restoring the system the following day. The expression of the will of an employee does not amount to authorization to leave the service, which requires the voluntary consent of both the employee and the employer.
The individual compensation scheme disclosed in the Journal is more than Euro650,000, mainly for highly qualified and paid employees and cannot be considered a universal standard. Full reports from Motor1 show that the programme can combine one-time payments, instalments, pensions and insurance, and that some staff members can retain their car purchase preferences after separation.
The BMW goal is to reduce the number of posts by about 800 by 2027.
Today AI cumulative financing over $200 million, Tencent, ants, etc.
According to Thunder, Today AI announced on October 21 that the team had cumulatively completed over $200 million in financing, with investments in IDG capital, Tencent, five-source capital, ant groups, Redwood China, voice-control units, acoustic investment and Shanghai-based capital.
Today AI was founded by Zhi Junmoto, founder of Teambition and former Doubao PC executive, and is based on personalized, active AI assistant. The product is already online in the country and can perform tasks such as calendar management, information research, work assistance and personal documentation.
According to intellectuals, Today supports iOS, Androids, Windows, Mac and Linux and can connect the enterprises WeChat, WeChat for reading, Baidu webpad, KIMshan documentation, Tencent meetings, etc. The team hopes to assist users before they are asked to do so through continuous interactive accumulation of user memory.
The Open AI model will be placed in Google Clouds, which enterprises can purchase using cloud consumption commitments
Open AI and Google Cloud announced a partnership to connect the front model and Codex Q42workflow to Model Gardner in Google Clouds, which is scheduled to be fully open in the coming weeks.
Enterprises can purchase these models through Google Cloud Marketplace at a cost that is offset by existing Google cloud consumption commitments. The two sides will combine model capabilities with Google Cloud access control, centralized data governance and infrastructure for software development and complex AI agent missions.
Open AI also indicated that its models would soon enter Google Antigravity and Gemini Agent, allowing enterprises to select models from these tools. Microsoft remains the main cloud partner of Open AI.
Yin Weida put AI in a container, DSX POD, 14 days fast.
At the GTC Berlin Congress, Ying Weida launched DSX POD: a full Vera Rubin AI plant, built by partners in modularized containers, completed integration and validation prior to release, scheduled to be delivered globally in early 2027.
According to Weedda, this option would reduce the field deployment time from 6 months to 14 days maximum, provided that the site ' s electrical, cooling, network, etc. are ready. The installation and testing of computing, network, wiring, hanger power and liquid cooling were moved to the manufacturing plant in parallel with site preparation.
Participating partners include Dale, Hewlett-Packard Enterprises, Union, Schneider Electrical and Vegetable Technologies. Wong In-hoon stated that the node at AI where the plant began to produce value was input model training or production reasoning, generation token, DSX POD wanted to compress the time between delivery and official operation.
NSB: Full implementation of AI+ to support corporate ownership of results
The CCP Central Office and the State Council ' s Office, publish, Opinions on Promoting the Deep Integration of Science, Technology and Industrial Innovation, propose a comprehensive implementation of the "Advisory Intelligence Plus" initiative to promote the full empowerment of artificial intelligence.
The Opinion calls for the acceleration of industrial Internet innovation to transform traditional industries with advanced technologies; the exploration of technological routes, applications and business models in frontier areas such as quantum technology, brain interfaces, smartness, and 6G; and the promotion of industrialization and commercialization.
In terms of the transformation of results, the targeted National Science and Technology Plan project for industrial applications is in principle led and led by the enterprise, encouraging universities and research institutes to license the results of science and technology to MSMEs on a first-use basis. At the same time, it proposes to promote the development of laws and regulations in the frontier areas of artificial intelligence, biotechnology, etc., and to explore ways of "sand box regulation".
It's almost $100 billion in new financing.
According to Ray, DeepSeek's new round of nearly $100 billion financing is almost complete, with a pre-investment valuation of about $500 billion. The report cites sources that both deep-seated and scavenging funds are involved in the current round and that Tencent continues to follow suit.
According to the Fund's progress announcement of the perfect world, Xinxiang has been raised to a total of $7.52 billion, and the company has pledged $50 million; the Elephant Fund has a total of $3.65 billion and the company has committed $100 million.
According to the information obtained by lightning, new investors in this round are dominated by industrial capital, including stone stream capital, central core concentration, CPE source peaks, King's Union capital and gili, and old shareholders in the Ninde era, Tencent, IDG capital, think capital and pick capital.
Codex Launching Permanent Dialogue cache
The Open AI Product Manager Tibo Sottiaux announced that API could now select 12 hours to alert cache; for log-in users of Codex and Work, the company would retain cache permanently. Users can also hit cache if they return to the old conversation or continue after a few hours of interruption.
Open AI Developments explains that the long-run AI agent will use the same context repeatedly in one day. The new option allows shared commands, tools and contexts to be retained for longer periods of time, even if the two requests are kept at a lower cache input rate.
Tibo updates this change as "day 16" and states that the resumption of the old session will not be subject to additional punishment.
Tesla's net profits declined by 43% in the third quarter. Free cash flow is positive.
Tesla published financial statements for the third quarter of 2026, collecting $28,466 million, which is essentially the same as the same period in the previous year; GAAP net profits attributable to ordinary shareholders were $787 million, a decrease of 43 per cent over the same period. The company delivered 486532 vehicles during the season, a 2 per cent decrease over the same period.
Several sets of data present changes in operations and expenditures during the season:
- Automobile operations received $25.60 billion, a decrease of 3 per cent over the same period; storage operations received $2881 million, a decrease of 16 per cent; and services and other operations received $6.05 billion, an increase of 33 per cent.
- GAAP ' s operating profits were $718 million, a decrease of 56 per cent over the same period, with an operating profit rate of 2.6 per cent; the amortized earnings per share were US$ 0.22 and the non-GAAP amortized earnings per share were US$ 0.48.
- Capital expenditure was $4,732 million, down from $5,789 million in the previous quarter; operating cash flows were $6,263 million and free cash flows were $1,531 million, which was reversed from negative values in the previous quarter.
According to the company, more than three quarters of new car deliveries in North America were accompanied by FSD subscriptions, and the number of active FSD subscriptions globally reached 1.74 million, with a calibre of one-time purchase and monthly subscriptions, excluding free trial tests. Tesla also stressed that the FSD still requires pilot active supervision. Cybercab joined the Robotaxi service in Austin on September 3, and Tesla Semi and Optimus started production in Nevada and Frimont, respectively.
IBM software business growth rebounded and the year-round free-flow expectations were updated
IBM announced performance for the third quarter of 2026, with a profit of $171.3 billion, an increase of approximately 5 per cent over the same period and an increase of 6 per cent at the fixed exchange rate. The net profit of the going concern GAAP was US$ 2,199 million, an increase of 26 per cent; the corresponding share was US$ 2.31 per share and the operating value of non-GAAP was US$ 2.95 per share.
The software business account received approximately $7.8 billion, an increase of 8 per cent over the same period. Of these, Red Hat grew by 11 per cent, data operations by 19 per cent and automated operations by 3 per cent. The consulting operations account for approximately $5.3 billion, which is essentially the same as during the same period of the previous year; the infrastructure operations account for approximately $3.7 billion, an increase of 4 per cent, but IBM Z revenue decreased by 20 per cent.
The company operated with cash flows of about $4 billion, free cash flows of about $3 billion and interest payments to shareholders of about $1.6 billion. The cumulative amount of free cash flows during the first three quarters was about $7.8 billion.
IBM is expected to maintain a 4 to 5 per cent increase in the year-round fixed exchange rate, raising the year-wide free-flow outlook to more than $1 billion per year. The Board also approved quarterly cash dividends of US$ 1.69 per unit on 10 November and 10 December for payment.
The faster the code is written, the greater the demand for testing: Blacksmith is valued at $1 billion.
According to Fortune, the code-testing infrastructure company Blakesmith completed a $66 million C round of financing, valued at $1 billion. Established in 2024, the company provides algorithms and software for code consolidation and pre-deployment automatic build-up and testing.
Blacksmith claims that more than 10,000 businesses use its platform, including Granola, Handshake, Supabese and Langchain. The number of certification missions per week increased from 5 million at the beginning of the year to 9 million at present, with annual recurrent income increasing 17 times over the same period.
The co-founder and CEO Aditya Jayaprakash says that the AI tool makes writing codes easier and bottlenecks turn to authentication codes. Blakesmith purchased its own server, deployed in the data centre and required duplicate preparation through software reuse testing.
This round of funds will be used primarily to purchase more computing power. Companies also face constraints in computing and memory supplies, as well as competition from model manufacturers to provide their own code certification services.
Peng Ho: Autopilot should skip L3. Safety is a question mark.
The President of the Board of Directors of Peng Motors, Ho Xiao Peng, said at the World Smart Net Motor Congress that automatic driving should skip L3. In his view, the L3 model transferred responsibility for security to the system, but the user experience was not optimized and the question mark was also needed for increased security.
In his judgement, L3 is essentially closer to the enhanced form of L2, and his preferred course of development is L2, L4, and the redefined L5. National regulatory systems should support the evolution of automatic driving to a higher level, provided that it is safe and manageable.
It was also suggested that the former L5 definition was primarily based on technical dimensions and was not sufficiently considered in terms of safety and controllability.
The future definition of L5 should also be adapted to the applicable scenario, as vehicles are passive autopilots that provide services under certain conditions, and robots may require 24 hours of active mission.
Turning to land landing times, he distinguished between technical and policy developments: At the technical level, he expected L4 to land in the second half of 2028; from a regulatory and policy point of view, he could judge time between 2028 and 2032.
Anthropic, launch Claude Duet, will report and continue in the backstage.
Anthropic publish Claude Duet, who makes Claude an individual AI agent who will contact the user on an active basis to process the mission. The current list of candidates for the product open research preview is available for subscriptions to Claude Pro and Max in the United States, providing access to independent iOS, Android App and claude.ai.
Duet is based on Claude Opus 5.5, following existing memory, skills and application connections of users. It consolidates important matters in connection applications every morning and continues to focus on change during the day and only sends messages when there is a need for processing or decision-making. Users can also entrust it with scheduling meetings, producing forecasts, developing dashboards or developing web applications, and will continue its work after leaving.
It uses user-connected tools in cloud-end computers, and the product page claims to support over 6000 connectors and plug-ins. Precise confirmation is required for sensitive operations such as sending messages, payments, etc.; Anthropic indicates that Duet does not directly see passwords and log-in signs, and also adds a special information protection sorter to limit the flow of sensitive information such as identity documents, credit cards, etc.
Research previews are included in Pro and Max packages, shared quotas are used with Claude, and back-office checks consume quotas; Team and Enterprise are not included. The official tip is that the product may still misread the conversation, determine the urgency of the error, or omit the news between the two inspections.
State Secretary 3149. Glory Win2 Series.
On October 21st, two mobile phones were opened in Hangzhou, publish Win2 and WIN RT2 for sale on the day of publish, at a premium of $4999 and $3149 respectively. It also brings WIN Turbo2, WIN tablet and WIN headphones.
- WIN2: Carried the sixth generation of Tyrone 8 super-high-to-high, equipped with a two-way fan stereo-net radiator system with a total wind volume of 1.19 CFM; incorporated 10100 mAh batteries to support 100W wired and 80W wireless charge. There are also 200 million pixels, USB 3.2, 3D ultrasound fingerprints and custom-defined electric competition Win keys.
- Win RT2: A fifth generation of Tyrannosaurus 8-to-tops, equipped with an Uranus fan in the same section as WIN2, increasing the battery to 12070 mAh. Both mobile phones use 2K 185Hz double-chip oasis screens, as well as LPDDR5X Ultra RAM and UFS 4.1 flash.
- Win Turbo2: A pre-sale has been opened and will open at 10:08 on October 30th, at a paid price of $2251.65.
- Glory WIN tablet: Using 8.4 inch 240Hz OLED screens to carry the fifth generation of Tyrannosaurus 8 to the upper stage, the wind cool-dispersion system and 10000 mAH batteries.
- Glory WIN headset: Supports 40 ms low delay and 55dB noise reduction and authentication through Hi-Res.
WIN2 will also launch a joint venture on Delta Action, opening an electrician appointment on October 30th, scheduled for November.
Flowtica got new financing, Magnetic Ink Screen Hardware, Verso Plan Goods
According to the Kryptonite report, the AI hardware company Flowtica has completed a new round of tens of millions of yuan in financing, with investments in Makoto capital, source capital, five-source capital and Clean Water Bay funds. Its magnetic inkscreen hardware, Verso, was pre-collected this week and the first mass products were scheduled to be shipped in December.
Verso can be placed on the back of a mobile phone, 6.5 mm thick and 78g heavy, equipped with constant ink screens, entity keys, the Type-C interface, four microphone arrays and bone transfer sensors. The screen is used to present the calendar, to-do and AI extracts information, and the side buttons support blind recording and call AI.
The founder, Ryan, explained in an interview that the voice was suitable for input commands and that the screen was easy to quickly scan the results. Based on the team ' s product design, minutes of meetings will be linked to the individual long-term memory to generate to-do dismantling, pre-session briefing and follow-up drafts; the company had previously launched a tape pen, Scribe, Verso, as its second-phase hardware.
Dyson pushes out the water-lighted straight-board clips and cools down with a condenser when he pulls his hair straight.
In mainland China, Dyson has introduced Corrale Cool Shine water-lighted straightboard clips, adding disassembly disassembly heat condensable condensers to the soft side of the plate, so that the hair can simultaneously cool as it passes through the hot plate plastic.
According to the product materials provided by Dyson, the fittings are covered by a sealed copper tube, with a small amount of internal water taking heat during evaporation and reducing the amount of heat remaining on the fuse. The flexible plate is packaged and co-ordinated, so that the hair is equilibrated and the pull is reduced.
The new machine provides 145 °C, 165 °C, 185 °C and 210 °C, class IV temperatures of approximately 15 seconds for preheating with a net mass of approximately 390 grams of accessories, and supports global utility voltage. According to Dyson, post-form scenic scenicity is up to 154 per cent compared to unformed hair.
The product was distributed in mainland China on October 20th, providing two colours: twilight and twilight.
Biotech Open Motubrain 2.0 Training Process, scheduled for half a month
The Biotechnology World Model team announced that Motubrain 2.0 training was launched on October 21st, and that it was planned to last for half a month to demonstrate the process of physical models from training, testing, to competencyover. The team indicated that Motubrain 2.0 was about to open and that no specific open source date had yet been published.
The Motubrain series is oriented towards robotic missions and studies the capability across missions, scenarios and robotics. This live broadcast covered four stages: the impact of the pre-training phase on performance and the zero sample capability for generalization; the post-training phase to improve the implementation capacity of specific tasks; the Agent mission implementation phase to explore the migration of experience to other tasks; and the back-to-back self-improvement (RSI) phase to examine whether models could assess and improve their own actions based on implementation feedback.
The team stated that the study was still in the exploratory phase and that the results were subject to subsequent experiments and evaluations. The official Live Lab page displays training losses, training steps, model calibration curves and 4 road camera portals, which include headphones, drawers, pick-up items and whiteboards.
Woo-ki publish H2-X "Kung Fu Legend" with a maximum duration of 8 hours
Yesterday, the publish human robot H2-X "Kung Fu Legend Guardian."
H2-X was described as providing services and guarding with robust performance for a maximum duration of approximately 8 hours and alerting users to the safe and friendly use of robots. The video cover contains the Unitree H2-X model.
As officially announced earlier, 2026 World Physical AI and Wootree Partners Ecoconference will be held on 22 October, and H2-X will be open the day before the Congress.
Card Q343 Launch F1 special, carbon-fibre-shareding with exclusive car colour
In collaboration with Formula 1, a special version of the Q3 43 Formula 1 camera was announced for sale from October 21st through the Qommen shop, the official network shop and professional retailers.
The special edition brought the racing material to the camera: the body covered the black Microveloour materials similar to the steering wheel, and the handle, fingertip and soft door buttons used German handmade carbon fibre components. The top cap, shot cap, and back belts are marked with F1 and cards, while the F1 red is used for soft-door buttons and the right-hand function key of the pointer.
The camera also added a new "Formula 1" Leeca Look to present the atmosphere of the track through greater contrast and saturation. Users can pass the colour style to the camera via Leica FOTOS App, preview it in the viewer and apply it directly to JPEG.
The core specifications continue the standard version: 60 million pixels are all painted CCOS, Maestro IV processor, APO-Summicron 43mm f/2 ASPH.shot and up to 8K video recording. A carbon fibre handle supports wireless charge and Arca-Swiss fast-assembly structures.
quantization can run a single RTX 5090
The development of The Jedi for Life, KRAFTON Open Self-Research Language Model Raon-1 Mini preview, focused on Korean-language dialogue, programming and AI agent missions. The total model parameter is 28.8 billion, with token active about 4.3 billion parameters per processing, and original support 262144 token context.
It uses a mix of attention structure: 18 out of 25 floors uses an improved RDA based on Kimi Delta Attention, and 7 floors of full attention; the team also uses Raon-mHC to simplify DeepSeek's overlink to reduce the training count.
According to official self-measurement, the SWE-Verified code has been repaired in 70.8, but the Terminal-Bench Lite is 47.0, with different mission performances. The model provides a version of NVFP4 quantization, which can be reasoned on a single RTX 5090 with 32GB VRAM and supports MLX on the llama.cpp and Apple chips.
The weights and components of the downloaded model are licensed by Apache 2.0 and require access; the official hosting of API and the experience service have other terms. This is still a preview version, with possible adjustments to specifications and test results.
DGX Spark to form 8 clusters with a unified memory of up to 1 TB
At the GTC Berlin Congress, Yin Weida updates local AI products and software: DGX Spark can now manually configure up to eight devices to form a cluster, using built-in ConectX-7 network interconnections, with a single memory of up to 1 TB.
The cluster is geared towards small development teams and allows multiple people to connect to the same model at the same time, running more LLM, longer context and several AI agents. The NVIDIA Sync Model Launcher plan was launched later this month to download suitable models and containers by single machine or cluster configuration and access OpenCode.
With regard to hardware, the 64GB version of DGX Spark will be launched from partner channels on 23 October. The higher-positioned DGX State has provided 748 GB consistent memory, and the Windows version was launched by partners in December to establish workflow between local models and manufacturing professional software.
The October software upgrade on-line also adds DGX Spark to Wi-Fi Assistant, which allows equipment to switch networks in an application mode without monitors.
Starbucks, in October 22nd, we're going to have a nationwide "consumption seating" that's going to cost over $300 million in upgrades.
Starbucks China will introduce "consumption seating" throughout the country from October 22nd, and the door shop will display the signs. The company indicated that it would provide a unified service guide to its door-to-door partners, that it would be implemented in accordance with the principle of friendly communication and that reasonable contingencies and special circumstances would be properly addressed, and that it would give priority to clients in need of help, such as pregnant women, persons with reduced mobility.
This door shop space rule was published in conjunction with the New Year Service Upgrading Plan. Starbucks China announced that it would invest more than $300 million in the 2027 fiscal year to increase peak hours and select the bar ' s manpower, upgrade its intellectual systems and core equipment, and renovate about 1,000 stores.
Nearly half of the country ' s stores will increase their capacity at peak hours, adding more than 200,000 hours of service per month and selecting more than 70,000 hours. A new smart mover POS with more than 70 functions will be on line in stages from January 2027.
The service guarantee also includes the free re-engineering of the customer ' s dissatisfied drink and the assistance of the owner ' s store in re-manufacturing a single store by means of a valid order.
The U.S. Wal-Mart has a separate file for AI, which can be limited to users and withdrawn at any time.
Wal-Mart announced the introduction of Agent Products, Agent Gateway and a special AI agent usage clause for AI to cover Wal-Mart and Sam's member stores. Users can choose their custom AI applications to buy goods and manage permission for each AI agent separately.
Agent Products plans to go online in the next month. Each AI agent will have a separate file in the customer account and the purchase order will be shown separately from the individual order; the user may request confirmation on a case-by-case basis, may set consumption limits and close the authorization at any time.
Agent Gateway, which has entered the test, connects AI agent to the shop system through the MCP protocol, allowing them to directly search for goods and complete purchases, reducing the number of cartons and errors caused by the simulation browser operation. Wal-Mart stated that the system could help AI agent to find commodities that users often buy without having to disclose to it the full history of shopping.
The clause prohibits AI agent from acting beyond the customer ' s intent, modifying the customer ' s means of payment or reading information not required for the task, and the violation of AI agent is prohibited.
Wal-Mart also indicated that the same commodity presented the same price to AI AI agent as the direct purchaser, and that the applicable discounts and fees would be indicated before payment.
Dec. 31, Sand dunes 3
Sand dunes 3 announced that it would be released in China on 31 December. IMAX Official Synchronizes publish set-up information, stating that the film is filmed using IMAX "special" and contains 71 minutes of exclusive painting with custom mixes.
The film, which is directed by Dennis Willennewa, is the final chapter of its three-part dunes and continues to revolve around the fate of Paul Estretti when he takes power. Timothy Charamé and Paul Zandaia, Cheney, went to the end of the story, and the return of the dead brought Paul a new mystery.
The film was created by the Legendary Film Industry and Warner Brothers Exploration Group and is scheduled to be released in North America on December 18.
The Happy Choir will be restarted. Disney+ and Hulu officially book a new play.
"Glee" will return with a new play. Disney+ and Hulu are officially booked for re-enactment and the producer is still 20th Television.
According to Hollywood Reporters, in August, the main author, Ryan Murphy, said he was considering restarting, and in September he revealed that a new version of the play had been written with original co-authors Brad Farchak and Ian Brennan. The announcement of the reservation has not yet been published for actors and production teams, and the lead agreement is still being finalized.
The original show was broadcast in Fox from 2009 to 2015, six seasons around a high school choir in Ohio. It combines popular songs, school growth and personality relationships, and has received six Emmy awards, which are now available in Hulu.
Tom Cruz will get the AFI.
Auto-fetched and translated by AI Atlas. All rights belong to the original author.