早报｜苹果回应iPhone Duo内屏开裂/Anthropic公开上市申请，冲击史上最大IPO/小鹏Robotaxi获批L4级示范运营
💵Anthropic 提交 IPO 申请，拟以「ANT」登陆纳斯达克
📱5499 元起，iPad mini 8 发布，仍无高刷
💬苹果回应 iPhone Duo 内屏开裂：少量设备受影响，可联系 AppleCare 换机
⚡OpenAI 发布 GPT-6.1 Terra，API 价格减半、电脑操作提速
🎨苹果正式推送 iOS 27.2：Apple Music 迎来黑白新设计、iPhone Duo 支持 Apple Pencil
✨华为畅享系列累计发货突破 1.6 亿台，畅享 100 将于 10 月 28 日发布
🚗小鹏悠游获批 L4 级示范运营，初期仍凭邀请码体验
🧠DeepSeek Flash 1017 登顶 OpenRouter 周用量榜，处理 33.6 万亿 Token
💬SK 海力士回应特别股息传闻：方案未定，拟于 28 日公布股东回报安排
🏢Kimi 为 K3.1 Pro 限流致歉并重置额度
💼奥迪中国明确品牌分工：一汽主攻四环，上汽聚焦 AUDI 字母标
🏛️大众中国获得 L3 路测牌照，已在合肥启动实路测试
🔬Anthropic 确认部分内部系统遭研究人员访问，客户对话与模型权重未受影响
💾Cadence 推出 Maestro 芯片设计模型，可在客户防火墙内运行
🪟微软向约 6 万名员工提供自愿离职方案，最迟 12 月初决定
💵Meta 启动约 350 亿美元股权融资，另拟推出 600 亿美元售股计划
🚀SpaceX 获约 400 亿美元 GPU 融资承诺，将租赁英伟达 Vera Rubin 设备
💡比尔·盖茨：AI 带来富足前，要先度过约 20 年的动荡
✨卢伟冰称 REDMI K100「同档无敌」，新配色雪岩白公布
🪄魏牌高山 8 HEV 上市，不插电油混版限时置换权益价 25.98 万元起
🤖无问芯穹与三体宇宙发布智子 Agent，可延续跨会话记忆
🔓腾讯混元开源 Hy-Nomas，可用文字和镜头逐段生成视频世界
🎬即梦上线短剧工作室，剧本、分镜与局部重拍可在同一平台完成
✨腾讯开源中英同声传译模型 Hy-SimulMT，支持边听边译与音色克隆
🧠老干妈否认「600 人裁至 48 人」，确认已应用 AI 质检
🤝ROG ×《影之刃零》首批联名硬件开售，武侠设计覆盖主机与外设
🛍️大众点评内测「生活家」：找店、打卡和互动也能攒成长分
💡《三星堆：未来往事》上映第四天票房约 119 万元，博纳影业股价跌停
🎬2026《鸿蒙星光盛典》定档 12 月 5 日
✨路阳监制电影《试探》定档 11 月 21 日，发布新预告
Anthropic 提交 IPO 申请，拟以「ANT」登陆纳斯达克
Anthropic 10 月 26 日向美国证监会公开提交 S-1 注册文件，计划在纳斯达克以「ANT」为代码上市。公司称，拟发行股数和价格区间尚未确定，注册文件也尚未生效。
招股书披露，今年上半年营收为 162.86 亿美元，去年同期为 11.63 亿美元；GAAP 净亏损为 807.48 亿美元。其中，可转债、衍生负债及未来股权简单协议的公允价值变动造成 723.88 亿美元损失，主要与公司估值上升有关，不能等同于同期现金消耗。
经营层面的亏损为 6.93 亿美元。剔除股权激励、非现金慈善捐赠等项目后，经调整经营利润为 16.18 亿美元，属于非 GAAP 口径。
文件同时提示，未来收入增长可能放缓，算力和基础设施扩张会继续带来成本压力；高度自主的 AI 系统若被滥用或缺少适当防护，也可能造成灾难性风险。公司创始人实体持有的特殊 F 类股份，将拥有一般股东表决事项 50.1% 的投票权。
🔗 相关阅读：突发！Anthropic 公开上市申请，史上最大 IPO 要来了
5499 元起，iPad mini 8 发布，仍无高刷
昨晚，苹果在官网低调上线新款 iPad mini，换用 8.5 英寸 OLED 超视网膜 XDR 显示屏与 A20 Pro 芯片，机身薄至 5.4 毫米，无线局域网版重 255 克。屏幕升级后仍不支持 120Hz ProMotion 高刷新率。
新机的主要变化包括：
- 显示屏边框收窄 24%，HDR 峰值亮度最高 1200 尼特，最低亮度 1 尼特；256GB 和 512GB 机型可选纳米纹理玻璃。
- A20 Pro 采用 2 纳米工艺，均配 7 核 GPU；128GB、256GB 版为 5 核 CPU 与单 16 核神经网络引擎，512GB 版升级为 6 核 CPU 和双 16 核神经网络引擎。
- 新增 Acoustic Display 音频系统，并支持 IP67 级抗水防尘；前置 1200 万像素 Center Stage 摄像头移到长边。N1 芯片支持 Wi-Fi 7、蓝牙 6 和 Thread，蜂窝版搭载 C2 基带。
- 支持快速充电，搭配另购的 35W 或更高功率兼容适配器，约 30 分钟可充至最高 50%。
新款提供午夜色与星光色，128GB 无线局域网版 5499 元起，蜂窝版 6799 元起，10 月 28 日接受订购，11 月 6 日发售。同步推出的科技织物智能双面夹售价 599 元，支持 Apple Pencil Pro 与 USB-C 版 Apple Pencil。
🔗 相关阅读：新 iPad mini 发布：5499 元起！应有尽有，除了高刷
苹果回应 iPhone Duo 内屏开裂：少量设备受影响，可联系 AppleCare 换机
苹果首次回应 iPhone Duo 内屏开裂投诉。在发给媒体的声明中，苹果表示已注意到少量 iPhone Duo 的内屏存在问题，遇到显示异常的用户应联系 AppleCare 更换设备。
我们已获悉少数 iPhone Duo 设备的内屏存在问题。如果用户的 iPhone Duo 显示屏出现问题，应联系 AppleCare 更换设备。
苹果还表示，绝大多数设备不受影响，问题机型通常在使用初期就会显现故障。
据苹果员工透露，公司尚未向零售门店和客服人员下发正式指引，但已有团队展开调查，部分工程师正设法回收受影响设备进行分析。
此前，部分用户在 Reddit 等论坛上传了内屏沿铰链中线开裂的照片。iFixit 拆解提出过部件可能向屏幕施加过大压力的推测。
OpenAI 发布 GPT-6.1 Terra，API 价格减半、电脑操作提速
OpenAI 发布 GPT-6.1 Terra，主打视觉理解、电脑操作和自然写作，已向 ChatGPT Work 与 Codex 的 Plus、Pro、Enterprise 和 Edu 用户开放，开发者可通过 API 调用。
OpenAI 表示，新模型标准 API 价格比 GPT-5.6 Terra 和 GPT-6.1 Sol 低 50%：每百万输入 Token 为 1 美元、缓存输入为 0.05 美元、输出为 5 美元。
在包含建筑图纸转 3D 模型等专业任务的 Agents」 Last Exam 测试中，Terra 在中等推理强度下比上一代高出 15 个百分点，成绩与 GPT-6 Astra 相当；官方称，完成电脑操作任务的速度超过 GPT-6 Astra 和 GPT-6.1 Sol 的 3 倍。
编程方面，Terra 在最高推理强度下的 Terminal-Bench 4.0 通过率为 50.3%，上一代公开成绩为 21.5%。OpenAI 仍将 Sol 和 Astra 列为最复杂编程任务的优先选择。
🔗 相关阅读：刚刚，GPT-6.1 Terra 骨折价发布，Agent 时代的打工模型来了
苹果正式推送 iOS 27.2：Apple Music 迎来黑白新设计、iPhone Duo 支持 Apple Pencil
今天凌晨，苹果正式推送了 iOS 27.2 更新，针对多项原生应用进行设计调整，并扩展了 Siri 的语言支持。
在视觉与界面设计上，Apple Music 图标由传统的红色调整为黑白设计，选中的标签页强调色也由红色替换为蓝色，同时推出了黑色背景的音乐小组件。
本次更新还为 iPhone Duo 增加了对 Apple Pencil（USB-C）的支持。
在功能层面，Siri AI 语言能力得到进一步扩展，新增支持法语、日语、韩语、葡萄牙语和西班牙语等 5 种语言。
其他原生应用与功能也迎来了相应改进。
- 信息 App 新增「建议的操作」功能，可根据对话内容自动提供创建日程或提醒等快捷方式；
- 播客 App 在主页中将「新单集」与「正在播放」进行了分离展示；
- Apple TV 增加了对多用户配置文件的支持；
- Group FaceTime 的双向拍摄功能扩展至群组通话，允许用户在通话中同时分享前后置镜头的画面（国行不可用）。
智能家居方面，兼容的智能门锁新增了「关闭后自动上锁」和「开门过久提醒」选项，HomePod 更新后也已支持「锁定控制」。
华为畅享系列累计发货突破 1.6 亿台，畅享 100 将于 10 月 28 日发布
华为董事、终端 BG CEO 何刚 10 月 26 日发文表示，华为畅享系列走过 11 年，累计发货量已突破 1.6 亿台。他将耐用可靠和续航持久列为用户选择这一系列的原因，并感谢用户长期使用。
新一代畅享 100 系列将于 10 月 28 日 14:30 发布。此次公布的 1.6 亿台是畅享整个系列的累计发货口径，覆盖多代产品，不是畅享 100 的订单量或销量。
小鹏悠游获批 L4 级示范运营，初期仍凭邀请码体验
小鹏 Robotaxi 取得广州市智能网联汽车应用示范运营资格，可在广州全域开展 L4 级示范运营。南都·湾财社 10 月 26 日从小鹏集团获悉，初期「小鹏悠游」仍将通过邀请码逐步开放体验，并稳步扩大服务范围。
这是在道路测试和员工内测之后，面向真实出行服务的一次准入推进。今年 5 月，小鹏首台 Robotaxi 量产车下线；7 月打通线上下单、自动接驾至送达的全链路；8 月取得广州主驾无安全员道路测试资质。10 月，业务定名「小鹏悠游 XPENG YOYO」，打车小程序同步上线。
小鹏悠游基于小鹏 GX 打造，搭载 4 颗自研图灵 AI 芯片和第二代 VLA 大模型。何小鹏此前表示，业务定位是向合作伙伴提供整车平台、自动驾驶软件和 AI 能力，由伙伴承担当地车队运营，收入来自硬件销售、技术服务与佣金分成。
DeepSeek Flash 1017 登顶 OpenRouter 周用量榜，处理 33.6 万亿 Token
OpenRouter 截至 10 月 25 日的最近一周用量榜显示，DeepSeek V4.1 Flash 1017 处理了 33.6 万亿 Token，环比增长 801%，位居第一。
榜单第二至第五名依次为 DeepSeek V4.1 Flash、MiniMax M3.1 Flash 免费版、GLM 5.3 Flash 和腾讯 Hy4 preview，处理量分别为 18 万亿、13.7 万亿、13.2 万亿和 8.97 万亿 Token。
这份榜单统计经 OpenRouter API 处理的输入与输出 Token，排除用户或应用选择保密的请求，同一模型的不同变体单独排名。它反映平台内的使用量，不涵盖全球各厂商直连服务，也不是准确率或推理能力榜。
OpenRouter 另按文本请求次数统计厂商份额：在 10 月 19 日开始的一周，DeepSeek 占比为 26.5%，同样排名第一。请求次数占比与 Token 总量是两项不同指标。
SK 海力士回应特别股息传闻：方案未定，拟于 28 日公布股东回报安排
据韩国 Edaily 报道，SK 海力士 10 月 26 日回应特别股息传闻称，公司正研究多种提升股东价值的回报方案，目前尚未确定；拟在董事会决议后，于 10 月 28 日发布第三季度业绩时披露新增安排。
此前有韩媒称，SK 海力士考虑利用出售日本存储厂商铠侠股权所得收益，派发约 16 万亿韩元特别股息。
SK 海力士此前已宣布 40 万亿韩元的股份回购及全额注销计划。Edaily 报道称，回购于 10 月 21 日提前完成，所购股份将全数注销。
Kimi 为 K3.1 Pro 限流致歉并重置额度
Kimi 昨日就 K3.1 Pro 订阅计划中出现的限流和不必要的额度消耗公开致歉，并宣布对所有套餐的订阅用户做一次性用量重置。
Kimi 官方在 X 上表示，这些问题「出在我们」，本应更早修复。补偿方案是：所有套餐的每位订阅用户恢复月度额度的 25%，相当于一周用量。
当前用量低于 25% 的，直接重置为零。Kimi 还称，已调整套餐参数，正在减少不必要的子智能体调用和输出长度，让同样的额度完成更多实际任务。
此次事件发生在 K3.1 Pro 发布两天后。该模型的升级重点之一是金融研究，Kimi 内部测试显示，其金融建模得分由 K3 的 50.4 升至 57.6，金融研究得分由 58.8 升至 70.4。
奥迪中国明确品牌分工：一汽主攻四环，上汽聚焦 AUDI 字母标
奥迪在中国的两套品牌体系有了更清晰的分工。10 月 26 日，据第一财经记者武子晔报道，奥迪分别与中国一汽、上汽集团签署战略协议，明确四环品牌与 AUDI 字母标品牌的差异化定位，以及「双品牌、双合作伙伴」渠道架构。
中国一汽将强化奥迪四环品牌在智能豪华新能源车型和燃油车型两个领域的布局；上汽奥迪则聚焦 AUDI 字母标品牌，专注新能源市场，开发面向中国消费者的下一代智能网联汽车。
过渡期内，上汽奥迪仍可在 2027 年底前正常销售四环产品。上汽奥迪旗下既有四环产品，以及 AUDI 字母标产品的售后服务均将保持平稳运营。
大众中国获得 L3 路测牌照，已在合肥启动实路测试
大众汽车集团（中国）宣布，已在合肥获得 L3 级自动驾驶道路测试牌照，并启动真实道路测试与数据采集。
大众中国 10 月 26 日发布的公告称，这次获牌是向 2027 年实现 L3 级自动驾驶技术就绪迈进的一步。本次公布的范围为道路测试，集团给出的 2027 年节点也是技术就绪目标。
10 月 28 日，首届大众汽车集团（中国）技术日将在合肥举行，集中展示最新技术进展。
Anthropic 确认部分内部系统遭研究人员访问，客户对话与模型权重未受影响
据《华盛顿邮报》10 月 26 日报道，网络安全初创公司 Cathedral 在 Anthropic 的外部安全测试计划下，利用第三方企业软件漏洞进入其内部系统。Cathedral 的报告草稿称，AI 智能体运行不到 12 小时，就找到了访问会议实录、薪酬与录用信息，以及部分讨论模型的 Slack 消息的路径。
Anthropic 确认，研究人员访问了销售、人力资源和 IT 系统中的有限部分，涉及部分客户与合作伙伴的商业联系信息，以及基本账户记录；客户发给模型的消息、模型回复、AI 系统代码和模型权重均未被访问。
公司称，监控系统在 Cathedral 报告前已发现部分活动，并于当天切断访问；目前没有证据显示其他人利用了这些漏洞。涉事企业软件被停用，供应商确认修复后才重新启用。
双方对测试过程仍有分歧。Anthropic 认为 Cathedral 夸大了国家安全影响，且未完整披露在系统内的活动；Cathedral 则称已遵循安全研究惯例，把数据访问限制在必要的最低程度。
Cadence 推出 Maestro 芯片设计模型，可在客户防火墙内运行
芯片设计软件公司 Cadence 发布 Maestro AI 模型组合，用于驱动旗下芯片与系统设计智能体。模型可部署在 Cadence 云、公共云，也可放在客户自己的数据中心和防火墙内运行。
Maestro 部分采用英伟达开放权重 Nemotron 模型，并结合 OpenShell 安全技术；客户可以用自己的芯片设计数据微调。Cadence 将模型、强化学习流程与 EDA 设计工具连接起来，依据工程结果持续优化设计决策。
公司称，初期设计流程测试较通用基础模型实现最高 30% 的质量提升、最高 5 倍 Token 使用效率；这些是其早期测试结果。首批模型计划于第四季度向部分早期客户开放。Cadence 高级副总裁 Paul Cunningham 向路透社表示，Maestro 同时考虑开放和封闭权重模型路线。
微软向约 6 万名员工提供自愿离职方案，最迟 12 月初决定
据 Business Insider 查看员工方案并获得微软确认，微软向约 6 万名员工提供自愿离职补偿，覆盖美国、加拿大、英国和印度部分团队。符合条件的人数不代表最终离职人数；员工可在 12 月初前决定是否接受，并于年底离开。
补偿按职级、地区和任职年限分别计算。该媒体查看的美国方案中，64 级及以下员工每任职半年可获一周基本工资，65 至 67 级每半年可获两周基本工资；67 级及以下标准补偿上限为 39 周基本工资。
方案还包括继续归属的股票和医疗保险，其中 67 级及以下员工的标准股票归属期按工龄分为 6 个月或 12 个月。微软发言人表示，目前没有再次推出同类自愿离职方案的计划。
Meta 启动约 350 亿美元股权融资，另拟推出 600 亿美元售股计划
Meta 10 月 26 日宣布启动两项公开发行，合计约 350 亿美元，包括约 225 亿美元 A 类普通股，以及代表强制可转换优先股的约 125 亿美元存托股份。公司称，这些安排将支持 AI 基础设施及「个人超级智能」投入。
另一项约 600 亿美元的按市价售股计划，拟在新设 C 类普通股开始交易后启动，预计不早于 2027 年第一季度。两部分基础规模合计约 950 亿美元；未来计划尚未签署协议，实际启动时间、规模和条款仍未确定。
英伟达、AMD 和博通各表达了购买至多 50 亿美元本次普通股的意向，但均不具约束力，实际购买额可能调整。
这笔融资也分阶段到账。A 类普通股发行采用远期销售协议，Meta 初期不会直接取得该部分售股款，而将在未来结算时获得资金。公司表示，将以经营现金流、投资级债务、战略合作和股权资本共同支持算力建设。
SpaceX 获约 400 亿美元 GPU 融资承诺，将租赁英伟达 Vera Rubin 设备
SpaceX 10 月 26 日公布，与阿波罗全球管理公司管理的实体达成约 400 亿美元 GPU 及相关设备主融资租赁的约束性承诺，用于扩建地面算力基础设施。设备将包括英伟达下一代 Vera Rubin 芯片。
这笔融资由阿波罗与高盛牵头，高盛担任独家结构设计代理；贝莱德、高盛资产管理旗下基金及账户，以及 KKR 提供顾问服务的资本账户也作出核心投资承诺。
交易仍须满足约定交割条件，预计在 2026 年底前完成，具体租赁安排预计随 GPU 和相关设备于 2027 年上半年交付后启动。
SpaceX 表示，新增算力将支持面向企业与消费者的 Grok，以及现有和未来的云服务客户；租赁融资也是其分散资金来源、加快算力部署计划的一部分。
💡 比尔·盖茨：AI 带来富足前，要先度过约 20 年的动荡
比尔·盖茨在 10 月 22 日播出的 Decoder 播客访谈中表示，AI 能够减少住房、食物和医疗等领域的短缺，但他更担心通往富足的约 20 年过渡期：其间会出现赢家、输家和大量冲击。
在盖茨看来，电话销售、客服和会计等白领岗位会先受到影响，人形机器人随后也可能改变体力劳动。工作同时提供收入、目标感、同事关系与解决问题的机会，单靠商品变便宜无法回答人们如何获得这些东西。
他以自己的慈善工作举例：即使组织资源消除疟疾曾是他的独特贡献，AI 也可能做得更好，把他打发去打匹克球。对希望作出独特贡献的年轻人来说，这种能力变化会带来压力。
盖茨提出，人类可以选择把育儿、烹饪、艺术和体育等领域保留给人来做，即使 AI 能做得更好；提前退休、缩短工作周也是可讨论的办法。这些安排需要社会作出政策选择。他认为，一部分人可能主动远离技术前沿，以保留互相帮助和作出独特贡献的空间。
他同时强调，20 年只是自己选取的一个相对稳妥的估计，实际时间可能更长或更短。如果不大幅放缓或停止 AI 发展，机器人将逐步承担建房、种植食物和手术等工作；人类需要决定如何安排富足时代的生活，而他当前的重点仍是如何度过这段转型期。
卢伟冰称 REDMI K100「同档无敌」，新配色雪岩白公布
REDMI K100 将于 10 月 27 日 19:00 发布。小米集团总裁卢伟冰 10 月 26 日发文预告，并称其「又是一款同档无敌的产品」。
REDMI 官方同日公布新配色「雪岩白」，机身背部采用宣纸般的细腻纹理，并已在各大电商平台开启预约。
魏牌高山 8 HEV 上市，不插电油混版限时置换权益价 25.98 万元起
魏牌高山 8 HEV 于 10 月 26 日上市，推出舒享版和臻享版两款车型，官方售价分别为 26.58 万元、28.58 万元；汽车之家向魏牌官方取得的上市信息将限时置换权益价列为 25.98 万元、27.98 万元。
新车采用 2.0T 发动机与前后双电机组成的油电混动系统，全系标配 Hi4 四驱，配备 3.64kWh 电池，无需外接充电。系统综合最大功率为 322kW，峰值扭矩 782N·m，零百加速 5.9 秒，WLTC 综合油耗为 6.92L/100km。
车身长 5280mm、轴距 3145mm，采用三排七座布局，延续纯平地板和二、三排共轨设计。座舱使用 Coffee OS 3.5 与高通 8295P 芯片，支持 HiCar、Carlink 和 CarPlay。辅助驾驶系统为 Coffee Pilot 3，配置 192 线激光雷达、27 个传感器和 Orin-Y 芯片。
无问芯穹与三体宇宙发布智子 Agent，可延续跨会话记忆
无问芯穹与三体宇宙于 10 月 25 日在北京举行的「2026 三体日·脱水浸泡嘉年华」上宣布战略合作，并发布首个合作产品「智子 Agent 1.0」。这款 AI 对话智能体以《三体》中的小宇宙管理员智子为角色锚点，支持多轮交流和跨会话记忆。
双方分工中，三体宇宙负责世界观设定、角色边界和人格特征；无问芯穹提供智能体开发、知识与记忆管理、推理服务等基础设施。产品基于 AgentWorks 平台构建，将世界观资料整理为可检索的知识库，通过分层提示词管理角色语气，并用 Infini Memory 延续不同会话之间的记忆。
据无问芯穹介绍，对话生成时，角色表达与世界观知识检索并行进行；初稿还要经过独立审校模块，排查人设偏离和设定越界，再把具体冲突送回修改。系统负责定位明确的事实偏差，角色表达是否贴合原作则由三体宇宙内容团队把关。
双方后续还将探索角色长期记忆与关系演进、多角色共享世界观叙事，以及多模态交互。
腾讯混元开源 Hy-Nomas，可用文字和镜头逐段生成视频世界
腾讯混元视觉生成团队开源交互式视频世界模型 Hy-Nomas，并提供在线演示。用户先描述场景，再用多轮文字或相机轨迹控制后续片段，逐步改变物体、风格、场景与角色动作。
模型按片段自回归生成，并保留贯穿生成过程的 KV 缓存，帮助后续画面延续前面的场景。交互界面支持在片段之间继续输入指令，也可以用键盘调整镜头；公众号「产品笔记」的实测展示了先生成厨房里的猫，再继续指定猫起身、伸懒腰和向前行走的过程。
官方仓库已公开训练、推理、演示和数据标注代码，模型权重同步上线 Hugging Face。开发者可用自有视频做 LoRA 微调；低显存配置支持单张 RTX 4090，通过 INT8 权重与缓存、滑动窗口历史及 CPU 卸载降低显存需求，卸载也会增加数据传输开销。项目采用 Apache 2.0 许可。
即梦上线短剧工作室，剧本、分镜与局部重拍可在同一平台完成
即梦 AI 10 月 26 日宣布上线「短剧工作室」，将剧本、美术设定、分镜、后期和成片重制放进同一套创作流程。网页端已提供上传剧本、剧本创作和重制转绘入口。
官方介绍，自研剧本模型 DreamScript 支持原创、改编、出海、解说等题材，并可解析 10 万字长篇文本；图片模型 Anycook 面向短剧角色与场景设计，分镜模型 DreamShot 则支持画布、故事板及 3D 导演台预演。
后期工具可对不满意的单个片段重新生成，分离人声与背景音，或拆分镜头。重制转绘可识别已有短剧的角色、场景和镜头结构，再按目标风格、画幅与市场生成不同版本。官方页面还提供智能预演、视频拉片和最高 4K 超分功能。
腾讯开源中英同声传译模型 Hy-SimulMT，支持边听边译与音色克隆
腾讯混元 10 月 26 日发布并开源 Hy-SimulMT 同声传译模型，支持中文与英文之间的双向语音翻译。模型可持续接收说话人的音频，边听边输出译文和语音，并通过实时音色克隆保留说话人的声音特征。腾讯 Hy 翻译 App 已提供体验入口。
开源项目包含 SimulInfer 推理服务和 SimulTrain 训练工具，模型提供翻译输出与原文、译文交错输出两种版本，并有 FP8、GGUF 等格式。开发者部署时，语音输出与音色克隆还需要配置 TTS 服务。
腾讯公布的 RealSI 长音频评测中，模型的 StreamLAAL 指标约为 3 秒，表示译文相对输入音频的流式延迟，并非首次语音输出耗时。按该评测的 GEMBA 翻译质量评分，其结果高于参与对比的开源模型及 5 个商业 API 中的 4 个。
团队还用 18 段约 6 分钟的中英文音频，与 6 位专业同传译员作了小规模比较；结果只适用于所用样本、参与者和测试设置，不能据此推断所有场景都可替代人工同传。
老干妈否认「600 人裁至 48 人」，确认已应用 AI 质检
老干妈回应了「AI 质检上线后，拣椒工人从 600 余人减至 48 人」的传闻。据凤凰网《风暴眼》10 月 26 日报道，其记者致电老干妈人力资源部门，工作人员明确表示，网传人员调整内容为假消息。
该工作人员确认，AI 质检技术已在生产基地上线应用，但企业从未有过对应规模的人员调整安排，并称「从来都没有用过这么多工人」。
ROG ×《影之刃零》首批联名硬件开售，武侠设计覆盖主机与外设
ROG 与灵游坊 S-GAME 合作推出的《影之刃零》首批联名硬件于 10 月 26 日 10:00 开售，覆盖迷你主机、显示器、键盘、耳机、鼠标和键帽包。
官方产品页将黑红配色、刀刃纹理、武器雕刻、角色剪影与绯红血线融入硬件外观。NUC 16 联名迷你主机的半透明前面板印有主角「魂」的剪影，支架可以拆卸，支持横放和竖放。
NUC 16 联名版可选配酷睿 Ultra 9 290HX 处理器及 RTX 5080 笔记本电脑 GPU，采用三风扇和一体式均热板，支持免工具升级内存与固态硬盘。游民星空报道，联名机箱将于 10 月 29 日单独开售；购买指定产品并完成官网登记，可参与数量有限的游戏兑换码赠送活动。
大众点评内测「生活家」：找店、打卡和互动也能攒成长分
大众点评正在部分城市内测「生活家」用户成长体系。界面新闻 10 月 26 日从多位用户处获悉，并经记者测试确认，合肥、郑州、福州等地部分用户已能在 App「我的」页面看到入口。
与原有主要依靠发布评价积分的 Lv 体系相比，「生活家」覆盖更多日常行为。评价、笔记、打卡、大众评审、地点标记与贡献信息均可获得成长分，找店、点赞、回复他人评价等互动也被纳入；计划强调足迹记录，以及帮助其他用户发现好去处。
新体系共设 1 至 18 级，权益除免费试、到店礼外，还包括打车券、券包等。长期保持评价和记录习惯的老用户，在新体系中的等级通常也较高。此次为部分城市灰测，尚非全量用户切换。
《三星堆：未来往事》上映第四天票房约 119 万元，博纳影业股价跌停
AI 深度参与制作的院线长片《三星堆：未来往事》10 月 23 日上映后，首个周末的市场表现偏冷。据《每日经济新闻》援引猫眼专业版数据，截至 10 月 26 日午间，这部 100 分钟影片的累计综合票房为 119.2 万元，24 日、25 日场均观影人数均约 2 人。
26 日午间页面显示，当日排片 8469 场，上座率约 0.1%。同一天，制作方博纳影业股价跌停，收盘跌幅为 10.06%。此前，公司曾于 9 月 28 日、10 月 9 日和 12 日三个交易日涨停。
博纳影业已在 10 月 12 日的股票交易异常波动公告中提示，该片票房收益存在不确定性。公司另一部 AI 院线电影《计划外：广寒行动》已进入制作阶段，截至该公告披露日尚未取得公映许可证，也未正式定档。
2026《鸿蒙星光盛典》定档 12 月 5 日
2026《鸿蒙星光盛典》将于 12 月 5 日在深圳湾体育中心「春茧」体育场举办。央视新闻于 10 月
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。
Morning Apple responding to iPhone Duo's internal screen break/Anthropic's public listing application, approved L4-level demonstration operation with impact on the largest IPO/Pipotaxi in history
Anthropic submits an IPO application to land in NASDAQ with ANT
iPadmini 8publish, still clean.
An iPhone Duo internal screen split: A small amount of equipment is affected and ApleCare is contacted for replacement
⚡Open AI publish GPT-6.1 Terra, API Halve Price, Computer Operation Speed
Apple Apple officially sends iOS 27.2: Apple Music for new design, iPhone Duo for Apple Pencil
Accumulative shipments from the Zhua series have broken by 160 million stations, 100 of which will be on October 28, publish.
The L4-level demonstration was approved for Peng-Yu, and the initial experience was based on the invitation code.
🧠DeepSeek Flash 1017 top of OpenRouter weekly scale, processing 33.6 trillion Token
💬SK Hercules responded to the ad hoc stock-taking rumour that the plan was not in place and that a shareholder return arrangement would be published on 28th
Kimi apologizes for K3.1 Pro Z104Z and reset quota
Audi's China's clear-cut division of brands: one capital hit four rings, focus on AUDI
In China, China has obtained L3 license plate and has started a live road test at the co-fertilizer.
Anthropic confirms that some internal systems were accessed by researchers and that client dialogue and model weights were not affected
💾Cadence rolls out the Maestro chip design model, which can run inside the client firewall
A voluntary separation programme for approximately 60,000 staff was provided by Microsoft and decided by early December at the latest
💵Meta initiates approximately $35 billion in equity financing and proposes an additional $60 billion sales equity plan
SpaceX received approximately $40 billion in GPU financing commitments to lease British Vera Rubin equipment
Bill Gates: AI will have to survive the turmoil of about 20 years before it brings wealth.
REDMI K100, New Colored Snow White.
8 HEV listed without plugging in the oil mix limited-time swap rate 25,98 million.
publish Wise Son of the Tristroscope.
🔓 Tencent hybrid open source Hy-Nomas, generate video worlds by paragraph using text and shot
Z Jimeng on-line short play studio, scripts, storyboard and local reshoots can be done on the same platform.
Hy-SimulMT to support interpretation and voice cloning
Old Godmother denied "600 to 48" and confirmed that the AI examination was applied
The first group of hardware to be sold by the roG x "The Shadow" is designed to cover the mainframe and the outside.
"Living home" in public opinion: looking for a shop, playing cards, and interacting can also grow. min
"The Three Stars: The Future" shows the fourth-day ticket room.
December 5th.
On November 21, publish's new announcement
Anthropic submits an IPO application to land in NASDAQ with ANT
Anthropic publicly submitted S-1 registration documents to the US SEC on October 26, planning to be listed in NASDAQ under the code ANT. According to the company, the number of shares to be issued and the price range have not yet been determined and the registration documents have not yet entered into force.
The equity book disclosed that in the first half of the year, $168.86 billion had been collected, compared with $1,163 million for the same period last year; the net loss of GAAP was $80,748 million. Of this amount, $72,388 million was lost as a result of changes in the fair value of reversible debt, derivative liabilities and simple future equity agreements, mainly related to the rise in corporate valuations and could not be equated with cash consumption over the same period.
The loss at the operational level was $693 million. After the removal of equity incentives, non-cash charitable donations, the adjusted operating profit was US$ 1,618 million, which was not GAAP calibre.
At the same time, the paper suggests that future income growth may slow and that the expansion of computing power and infrastructure will continue to put cost pressure; and that a highly autonomous AI system that is misused or lacking adequate protection may also pose catastrophic risks. The special F shares held by the company ' s founder entity will have the right to vote in general shareholder voting matters of 50.1 per cent.
🔗 Read about: Quick! Anthropic public application, the biggest IPO ever coming.
5499, iPad Mini 8publish, still clean.
Last night, apples were on the net low-key new iPad mini, with 8.5-inch OLD ultra-retinate XDR display and A20 Pro chip, thin to 5.4 mm, and wireless local area network version with 255 grams. 120Hz Promotion is not supported after the screen is upgraded.
The main changes in the new aircraft include:
- Displays a narrow frame of 24 per cent, with a peak HDR of up to 1,200 Nit, with a minimum brightness of 1 Nit; 256 GB and 512 GB machine types with an optional nano texture glass.
- A20 Pro uses a 2 nanoprocess with 7 nuclear GPU; 128GB, 256GB editions with 5 nuclear CPU and single 16 nuclear neural network engines, 512GB version with 6 nuclear CPU and double 16 nuclear nervous network engines.
- Add the Acoustic Display audio system and support the IP67 class for water protection against dust; advance 12 million pixels of the Center Stage camera to the long edge. The N1 chip supports Wi-Fi 7, Bluetooth 6 and Thread, the cellular version of which carries C2 base belts.
- Supports rapid charging, with an additional 35 W or higher power compatible appliance, of approximately 30 minutes up to 50%.
The new provision provides midnight colour and starlight, 128GB Wi-Fi version 5499, cellular version 6799, order on 28 October and sale on 6 November. Synchronizes the launch of a technology fabric smart double-sided sale price of $599 to support Apple Pencil Pro and USB-C versions of Apple Pencil.
Read about it: New iPad Mini publish: 5499! Everything, except a high brush.
Apple Respond iPhone Duo Internal Screen Break: A small amount of equipment is affected and ApleCare is contacted for replacement
Apple responded for the first time to an iPhone Duo break-in complaint. In a statement to the media, Apple indicated that it had noticed a problem with the inner screen of a small number of iPhone Duo, and that AppleCare should be contacted by users who showed anomalies to replace the equipment.
We have been informed that there is a problem with the inner screen of a few iPhone Duo devices. If there is a problem with the iPhone Duo screen of the user, contact AppleCare to replace the device.
Apple also indicated that the vast majority of the equipment was not affected and that problem models usually showed failure at the early stages of use.
According to Apple employees, the company has not issued formal guidance to retail outlets and customers, but teams have conducted investigations and some engineers are trying to recover the equipment affected for analysis.
Prior to this, some users sent photographs of the inner screen cracking along the middle line of the chain in forums such as Reddit. iFixit disassembly suggests that parts may exert excessive pressure on the screen.
Open AI publish GPT-6.1 Terra, API halved prices, speed up computer operations
Open AI publish GPT-6.1 Tera, which deals with visual understanding, computer operations and natural writing, has been made available to users of ChatGPT Work and Codex Plus, Pro, Enterprise and Edu, who can be called through API.
Open AI states that the new model standard API price is 50% lower than GPT-5.6 Terra and GPT-6.1 Sol: $1 per million input Token, $0.05 per cache and $5 per million input.
In the "Last Exam" test, which included specialized tasks such as building drawings and 3D models, Terra had 15 percentage points higher than the previous generation under the medium reasoning intensity and had the same grades as GPT-6 Astra; officially, the speed of completing computer operations exceeded 3 times that of GPT-6 Astra and GPT-6.1 Sol.
In terms of programming, Terra had a 50.3 per cent pass rate of Terminal-Bench 4.0 at the highest level of reasoning and 21.5 per cent for the previous generation. Open AI still lists Sol and Astra as the most complex programming priorities.
Read about it: Just now, GPT-6.1 Terra's fracture price publish, Agent's working model.
Apples officially send iOS 27:Apple Music
In the early morning hours, apples officially sent iOS 27.2 updates, design adjustments for many native applications, and extended Siri ' s language support.
In visual and interface design, the Apple Music icon was adapted from the traditional red to black and white, and the selected tab page highlighted the colour was replaced by red to blue, while the black background music group was launched.
This update also increased support for Apple Pencil (USB-C) for iPhone Duo.
At the functional level, the Siri AI language skills have been further expanded, with five new languages supporting French, Japanese, Korean, Portuguese and Spanish.
Corresponding improvements have been made to other primary applications and functions.
- Information App add a "recommended operation" function, which can automatically provide a shortcut such as a creation calendar or a reminder based on the content of the dialogue;
- Podcast App shows "New Singles" and "Playing" on the home page;
- Apple TV increased support for multi-user profiles;
- The two-way filming function of Group FaceTime was extended to group calls, allowing users to share both images of the front- and back-to-back shot (national lines are not available).
In terms of intelligent homes, compatible smart door locks have added the options of "Auto-Locking" and "Alarming Opens Too Long" and HomePod has been updated to support "Lock Control".
Accumulated shipments of China's Queen's Treasure series have broken 160 million, 100 will be released on October 28th.
In a letter dated October 26th, the Director, Director and Director of the BG CEO stated that China had travelled 11 years to enjoy the series and that cumulative shipments had exceeded 160 million. He listed the reliability of durability and durability as among the reasons for user selection and thanked users for their long-term use.
A new generation of 100 series will take place on October 28, 14:30 publish. This release of 160 million units is the cumulative calibre of the full series, covering multiple generations of products and not 100 orders or sales.
Peng-Yu was approved for the L4-level demonstration operation, and the initial experience was based on the invitation code.
Peng Robotaxi has acquired the competence of the Guangzhou Networked Smart Automobile Applications Model Operator, which can operate at the L4 level throughout Guangzhou. On October 26th, Nandu Bay Bank learned from the Peng Group that the initial Peng Run will be gradually open to experience through the invitation code and will steadily expand the range of services.
This was an access advance for real travel services following road tests and staff internal surveys. In May of this year, Rhobotaxi, the first Peng company, was taken off-line; July was used to order off-line and automatically connect to the full chain; and August was given the master of Guangzhou's securityless road test. In October, the business became known as "XPENG Yoyo" and the cabling applet went online simultaneously.
Peng is based on Peng GX and carries four self-study Turing AI chips and the second generation VLA LLM. He Xiao Peng previously stated that operational positioning was to provide partners with a whole-car platform, auto-driving software and AI capacity, with partners assuming local fleet operations and income from hardware sales, technical services and commissioning.
DeepSeek Flash 1017 top-up OpenRouter weekly scale, handling 33.6 trillion Token
OpenRouter's last week to October 25th showed that DeepSeek V4.1 Flash 1017 processed 33.6 trillion Token, an 801 per cent increase in the ring, and ranked first.
The second to fifth places on the list are DeepSeek V4.1 Flash, Mini Max M3.1 Flash, GM 5.3 Flash and Tencent Hy4 preview, respectively, with 18 trillion, 13.7 trillion, 13.2 trillion and 897 trillion Token.
The list shows the input and output processed through OpenRouter API, token, which excludes the user or application of a request to choose confidentiality, with different variants of the same model ranked separately. It reflects the amount of use within the platform, does not cover the direct services of global manufacturers, and is not an accurate rate or reasoning competency list.
OpenRouter also calculates the share of the manufacturer by the number of text requests: DeepSeek is 26.5% in the week starting October 19, also ranked first. The ratio of requests to Token total is two different indicators.
SK Hercules responded to the ad hoc stock-taking rumour that the plan was not in place and that a shareholder return arrangement would be announced on 28th
According to Edaily, Korea, in response to a special stock-sharing rumor on 26 October, SK Hercules reported that the company was studying a variety of return options for increasing shareholder value, which had not yet been determined; it was proposed that the new arrangements would be disclosed in the third quarter of publish performance on 28 October, following a Board resolution.
Formerly, according to the Korean media, SK Hercules is considering issuing about 16 trillion won of special dividends from the proceeds of the sale of shares of the Japanese store manufacturer Zhang.
SK Hercules had previously announced a repurchase of 40 trillion won and a full write-off plan. Edaily reported that the buy-back was completed ahead of schedule on 21 October and that the purchase of the shares would be fully cancelled.
Kimi apologizes for K3.1 Pro Z104Z and reset quota
Kimi publicly apologized yesterday for the rate limit and unnecessary quota consumption in the K3.1 Pro subscription schedule and announced a one-off replacement of all subscriptions to the package.
Official Kimi said on X that these problems were "we" and should have been repaired earlier. The compensation scheme is 25 per cent of the monthly quota, equivalent to one week of usage, for each subscriber to all packages.
Directly reset to zero if the current usage is less than 25%. Kimi also stated that the package parameters had been adjusted to reduce the unnecessary AI agent call and output length to allow the same quota to perform more actual tasks.
The incident occurred two days after K3.1 Pro Z94Z. One of the priorities for upgrading the model was financial research, which showed that its financial modelling score had risen from 50.4 to 57.6 for K3, and from 58.8 to 70.4 for financial research.
Audi China has a clear division of brands: a capital hit four rings and focus on AUDI
There is a clearer division of labour between the two brand systems in China. On October 26, Audi signed a strategic agreement with China's first financial reporter, Wu Zhu, to clarify the difference between the four-ring brands and the Audi alphabet, as well as the "two brands, two partners" channel structure.
China’s first steam will strengthen the layout of the Audi Four Circle brands in the areas of smart, luxurious, new energy and fuel-car types; and Upper Odie will focus on the Audi alphabet brand, focusing on the new energy market and developing the next generation of smart Internet-connected vehicles for Chinese consumers.
During the transition period, the upper Audi could still sell the four-ring product normally by 2027. The four-ring product in the upper Audi flag, as well as after-sale services for the AUDI letter-marked product, will be operating smoothly.
China's got L3 license plate, and it's already running a real road test at the Fats.
The Volkswagen Motors Group (China) announced that it had obtained L3-class auto-driving road test plates and initiated real road testing and data collection.
The announcement of publish of October 26th in China stated that the bid was a step towards achieving L3 self-pilot readiness by 2027. The scope of this announcement is road testing, and the 2027 node given by the group is also a technical target.
On 28 October, the first Volkswagen Motors Group (China) technical day will be held in Fats, focusing on the latest technological developments.
Anthropic confirms that some internal systems were accessed by researchers and that client dialogue and model weights were not affected
According to the Washington Post, on October 26, the cyber-security start-up company Catedral, under the Anthropic external security testing programme, uses third-party enterprise software loopholes to access its internal systems. According to Catedral ' s draft report, AI AI agent was running for less than 12 hours and found a path to accessing the minutes of the meeting, the pay and employment information, and the Slack message that partially discussed the model.
Anthropic confirmed that researchers had visited a limited part of the marketing, human resources and IT systems, involving business contact information from some clients and partners, as well as basic account records; none of the information, model responses, AI system codes or model weights that clients sent to models had been accessed.
According to the company, the monitoring system had detected some of its activities prior to Catedral ' s report and had cut off access on the same day; there was no evidence that others had exploited those loopholes. The enterprise software was discontinued and the vendor confirmed that it had been repaired before re-enactment.
The testing process remains divided between the parties. According to Anthropic, Catedral exaggerates national security implications and does not fully disclose activities within the system, while Catedral states that it has followed security research practices and limited data access to the minimum necessary.
Cadence rolls out the Maestro chip design model, which can be run within the client firewall
Chip Design Software Company Cadence publish Maestro AI model combination to drive AI agent under the flag and system design. Models can be deployed in Cadence clouds, common clouds, and can also be placed in client data centres and firewalls.
Maestro partially uses the Newmotron model, which combines OpenShell security techniques; customers can design fine-tuning data using their own chips. Cadence links models, enhanced learning processes and EDA design tools to continuously optimize design decisions based on engineering results.
According to the company, the initial design process test achieved a maximum of 30 per cent quality improvement and a maximum of five times the efficiency of the use of Token as compared to the generic base model; these were the results of its early tests. The first models are planned to be open to some early clients in the fourth quarter. Paul Cunningham, Senior Vice President of Cadence, told Reuters that Maestro was also considering an open and closed weight model route.
Microsoft offers a voluntary separation programme to approximately 60,000 staff members, to be decided by early December at the latest
According to Business Insider, who looks at the employee programme and obtains Microsoft confirmation, Microsoft provides voluntary severance pay to some 60,000 employees, covering parts of the United States, Canada, the United Kingdom and India. The number of eligible persons does not represent the final number of separations; the staff member may decide to accept them by early December and leave at the end of the year.
Compensation is calculated on the basis of grade, area and length of service. In the U.S. scheme, which is viewed by the media, employees in grades 64 and below receive a one-week basic salary for half-year service, and two-week basic salary for grades 65 to 67; the standard compensation ceiling for grades 67 and below is 39 weeks basic salary.
The programme also includes stocks and health insurance that continue to be attributed, of which the standard period of equity attribution for employees at grade 67 and below is divided into 6 or 12 months of service. The Microsoft spokesperson indicated that there were no plans to reintroduce a similar voluntary separation programme.
Meta launched some $35 billion in equity financing and proposed an additional $60 billion sales equity plan
Meta announced the launch of two public releases on 26 October, totalling approximately $35 billion, including approximately $22.5 billion in general category A shares and approximately $12.5 billion in deposit shares representing mandatory transferable priority shares. According to the company, these arrangements will support AI infrastructure and "personal super smart" inputs.
Another approximately $60 billion market-based equity plan, to be launched after the opening of the new general category C equity, is expected to be in place no earlier than the first quarter of 2027. The combined base size of the two components is approximately $95 billion; the future plan has not yet signed an agreement and the actual start-up time, size and terms remain undetermined.
In Weida, AMD and Chase each expressed their intention to purchase up to $5 billion of this common stock, but none of them are binding and the actual purchases may be adjusted.
This financing was also phased in. A. Generalized equity issuances are based on forward sales agreements, and Meta will not receive that portion of the share directly at an early stage, but will be funded at a future settlement. The company indicated that it would support computing together with operating cash flows, investment-level debt, strategic cooperation and equity capital.
SpaceX received approximately $40 billion in GPU financing commitments to lease British Vera Rubin equipment
SpaceX was released on 26 October, with a binding commitment of approximately $40 billion in the GFU and related equipment master finance leases with the entities managed by the Apollo Global Management Corporation for the expansion of terrestrial computing infrastructure. The equipment will include the next generation of Vera Rubin chips.
The financing was led by Apollo and Goldman Sachs, who acted as sole structural design agent; Beled, Goldman Sachs asset management funds and accounts, and KKR capital accounts providing consultancy services, also made core investment commitments.
The transaction remains subject to the agreed delivery conditions and is expected to be completed by the end of 2026, and the specific lease arrangement is expected to commence after the delivery of GPU and related equipment in the first half of 2027.
SpaceX states that the new calculator will support the company and consumer-oriented Grok, as well as existing and future cloud service clients; leasing finance is also part of its decentralized funding and faster computing deployment plan.
💡 Bill Gates: AI will have to spend about 20 years of turmoil before it brings wealth.
Bill Gates, in his interview with Decoder podcast on October 22, stated that AI could reduce shortages in housing, food and health care, but he was more concerned about the transition to wealth of about 20 years: during which there would be winners, losers and massive shocks.
In Gates ' view, white-collar positions such as telephone sales, passenger uniforms and accountants may be affected first, and human robots may subsequently change physical labour. Work provides opportunities for income, a sense of purpose, co-worker relations and problem solving, and cheaper goods alone cannot answer how people get them.
He gave an example of his charity work: even if organizing resources to eradicate malaria had been his unique contribution, AI could have done better by sending him to play Pike. For young people who wish to make a unique contribution, this change of capacity can put pressure on them.
Gates suggested that human beings could choose to retain the fields of childcare, cooking, art and sports, even if AI could do better; early retirement and shorter workweeks could be discussed. These arrangements require social policy choices. In his view, a segment of the population might take the initiative to move away from the technological frontier in order to preserve space for mutual assistance and unique contributions.
At the same time, he stressed that 20 years was a relatively conservative estimate of his own choosing and that the actual time might be longer or shorter. If AI does not slow down or stop development, robots will gradually take on the task of building houses, growing food and operating; human beings will need to decide how to organize their lives in a rich age, and his current focus remains on how to survive this transition.
Lou Wei Bing called REDMI K100 "Invincible", new coloured snow rock white.
REDMI K100 will be on October 27th at 1900 publish. On October 26th, the president of the Mee Group, Lou Wei Bing, wrote an article in which he called it "an invincible product of the same class."
On the same day, the official REDMI announced a new colored snowstone white, with paper-like fine texture on the back of the aircraft and an appointment has been opened at major electric power stations.
Wei Takayama 8 HEV listed, limited-time replacement equity price without plugging in electric oil, 25.98 million.
On October 26th, HeV Wei was put on the market and introduced two car models, for both enjoyable and enjoyable, at an official sale price of $2.65 million and $285 million respectively; the official listings of the Motel House to Wei listed the limited-time swap price as 25.98 million and $279 million.
The new vehicle uses an oil-electric mixing system consisting of 2.0T engines and back-to-back double generators, which is fully fitted with Hi4 and equipped with 3.64kWh batteries without external charge. The maximum power of the integrated system is 322kW, peak 782N.m, zero-hundred-acceleration 5.9 seconds, WLTC combined fuel consumption is 6.92L/100km.
The size of the vehicle is 5280 mm and the length of the axle is 3145 mm, with three rows of seven layouts to continue the pure flat floor and two or three rows of orbit. The module uses Coffee OS 3.5 with the Chase 8295P chip to support Hicar, Carlink and CarPlay. The auxiliary driving system is Coffee Pilot 3, equipped with 192 laser radars, 27 sensors and Orin-Y chips.
publish Wise Son of the Triple Space.
publish announced strategic collaboration with the Triple Cosmos on October 25 at the "2026 Triangular Day Dehydrated Absorption Carnival" in Beijing and the first collaborative product of publish, "Agent 1.0". This AI Dialogue AI agent supports multi-wheel communication and cross-conference memory with the role anchor of the little cosmonaut, the brain of the tribody.
In the division of labour, the three universes are responsible for worldview setting, role boundaries and personality characteristics; the no-question dome provides infrastructure such as AI agent development, knowledge and memory management, and reasoning services. The product is based on the AgentWorks platform, organizes worldviews into a searchable knowledge base, manages role tone through layered prompt, and sustains memory between sessions with Infoini Memory.
According to the no-question core presentation, role expression takes place in parallel with worldview knowledge retrieval when dialogue is generated; the first draft is also subject to an independent revision module that screens people for deviations and defaults across borders before returning specific conflicts to change. The system is responsible for clearly located factual deviations, and the role expression is closed by a three-body cosmic content team.
The parties will then explore the evolution of the role long-term memory and the relationship, the multi-player sharing of worldview narratives and the multi-modular interaction.
Tencent hybrid open source Hy-Nomas, generate the video world by text and shot by paragraph
The Tencent hybrid visual generation team open source interactive video world model Hy-Nomas and an online demonstration. The user first describes the scene, then controls the follow-up session with multiple rounds of text or camera tracks, gradually changing the object, style, scene and role.
The model is self-repeated and maintains KV cache, which runs through the generation process, to help follow-up images to the scene. The interface supports the continued input of commands between clips, and can also adjust shot by using keyboards; the WeChat Official Account "Product Note" test shows how Mr. Kim became a cat in the kitchen, then continues to assign cats to rise, stretch and walk forward.
The official warehouse has been publicly trained, reasoned, demonstrated and coded for data labelling, and model weights are being synchronized up to Hugging Face. Developers can use their own video for LoRA fine tunes; low VRAM configuration supports single RTX 4090, lowering the demand for VRAM through INT8 weights and cache, slide window history and CPU offloading, which also increases data transfer costs. The project uses the Apache 2.0 licence.
Jimeng on-line short play studio, script, storyboard and local re-shoot can be done on the same platform.
On October 26th, Jimeng AI announced that it would go online as a "short play studio" and place the script, art design, storyboard, post-production and the pieces into the same creation process. The portal has been made available for uploading scripts, creating scripts and rewriting.
Officially, DreamScript supports originals, adaptations, out-of-the-sea, interpretation, etc., and can decipher 100,000 words of long text; a photographic model, Anycook, aims at short play role and scene design, and a storyboard model, DreamShot, supports canvass, storyboards, and 3D director's preview.
The post-production tool recreates unsatisfied individual segments, separates the person ' s voice from the background, or removes the head of storyboard. Reprogramming allows for the identification of role, scene and shot structures that already exist, with different versions by target style, painting and market. The official page also provides smart previews, video clips and up to 4K hyperpoints.
Tencent Open-source Central-English simultaneous interpretation model Hy-SimulMT to support interpretation and voice cloning
On October 26, publish and open source Hy-SimulMT, a two-way voice translation between Chinese and English is supported. The model is sustainable in receiving the speaker ' s audio, output the translation and voice while listening, and preserves the voice of the speaker through real-time voice cloning. Tencent Hy Translator App has provided access to experience.
The open source project consists of the SimulInfer reasoning service and the SimulTrain training tool. Models provide translation output to the original language, staggered translation output, and formats such as FP8 and GGUF. When the developers deploy, voice output and voice cloning also require the configuration of TTS services.
In the RealSI long-sound assessment published by Tencent, the model 's StreamLAAL indicator is about 3 seconds, which means that the translation is relatively slow to enter the flow of the audio and is not time-consuming for the first voice output. The results were rated by the quality of the GEMBA translation as higher than the open source model and 4 of the 5 commercial APIs involved in the comparison.
The team also made a small comparison with six professional simultaneous interpreters using a medium English sound of about 18 minutes; the results were applied only to the sample, participants and test settings used, on the basis of which it could not be assumed that all scenes could be a substitute for artificial assimilation.
Godmother denied "600 to 48" and confirmed the application of AI tests
"At the end of the inspection, the pepper picker fell from 600 to 48 people." According to Phoenix Network Storm Eye, on 26 October, its journalist called Drymother ' s Human Resources Department, and staff members made it clear that the Internet passers-by was false.
The staff member confirmed that AI quality control technology had been applied online at the production base, but that the enterprise had never had staff adjustment arrangements of the same size and stated that “there had never been so many workers used”.
The first group of hardware to be sold by the ROG x Shadow Blade zero.
The first joint hardware of Shadow Wrests Zero, launched in partnership with the Spirit Fair S-GAME, was opened on October 26th at 10:00, covering mini-hosts, monitors, keyboards, headsets, mouse and caps.
The official product page integrates black and red colours, blade texture, weapon sculptures, role clippings and red bloodlines into the hardware look. The semi-transparent front page of the U.N.U.S. mini-host 16 has a clipping of the hero 's soul' that can be removed from the shelf and support for a wide and vertical presence.
The NUC 16 joint version of the Ultra 9 290HX processor and the RTX 5080 laptop GPU, using three fan and one-size-size-size thermal panels, supports tool-free upgrades of memory and solid-state hard drives. Travellers reported that joint aircraft boxes would be sold separately on 29 October and that the purchase of specified products and the completion of the official network registration would allow participation in a limited number of game conversions.
A popular review of the "life family": looking for shops, cards and interactions can also grow. min
The popular opinion is taking stock of the "liveer" user growth system in some cities. The interface news was received from multiple users on October 26, and it was confirmed by a press test that some users in the fields of Fat Man, Zhengzhou and Fukuzhou were able to see the entrance on the App "My" page.
The "living family" covers more everyday behaviour than the old Lv system, which relied mainly on publish to evaluate points. Evaluations, notes, cards, mass reviews, site markings and contribution information are all given points of growth, as are interactions such as shop searches, praises, responses to others ' evaluations; the plan emphasizes footprint records and helps other users find good locations.
The new system has a total of 1 to 18 classes, and benefits include, in addition to free tests and shop visits, tickets, vouchers, etc. Older users who have long-standing evaluation and record-keeping habits are usually also at a higher level in the new system. This is part of the urban ash survey, which is not fully user-to-user-to-user.
"The Three Stars: The Future" is showing a fourth-day ticket room about $11.9 million.
The first weekend market performance was cold following the launch of AI's deep-seated production of a long film, The Three Stars: The Future. According to the Daily Economic News, which cites data from the cat-eye professional edition, as of midday on October 26th, the cumulative total of this 100-minute film was 119.22 million yuan, with an average of approximately 2 people on 24th and 25th days.
26 The noon page of the day shows that the number of seats on the 846 series was approximately 0.1 per cent. On the same day, the producer's producer, Bobona, fell by 10.06 per cent in stock prices. This was preceded by three trading days, 28 September, 9 October and 12 October.
The Boone Shadow has suggested in the announcement of abnormal fluctuations in the stock exchange on October 12th that there is uncertainty about the return on the ticket room. Another of the company ' s AI-line films, Unplanned: Operation Pandemic, was in the production stage and had not been issued with a public release permit or a formal filing as of the disclosure date.
December 5th, 2026.
On December 5th, the Hong Mon Star Festival will be held at the Shenzhen Bay Sports Center's "Chun Coon" Stadium. The news is in October.
Auto-fetched and translated by AI Atlas. All rights belong to the original author.