早报｜15999起，苹果发布M6 MacBook Pro/张一鸣以6700亿元登顶胡润百富榜/电视新规：不得播放开机广告
💻苹果发布 M6 MacBook Pro 和 iMac
📉iPhone 18 Pro 上市三周在华销量增 29%，美国与日本下滑
🏛️有线电视与 IPTV 新规：不得播放开机广告，首页须设免费专区
📱IDC：中国手机 Q3 出货下滑 15.2%，折叠屏连续四季度下降
🏆张一鸣以 6700 亿元登顶胡润百富榜，梁文锋首次上榜排第七
🧠邓煜：AI 公司应公开证明思路，让用户决定数据是否用于训练
🔬研究：豆包引用抖音的权重升至 31.8%
⚠️大众在海外召回约 7.27 万辆电动车，部分车型充电上限暂降至 80%
🛑《宝可梦大集结》海外版将于明年 3 月停服，保留资料馆
💬红果否认「年收入 400 亿元」传闻，称大部分收入用于内容投入
🏠小米武汉首个员工住宅开盘，297 套房限内部员工购买
🍎AMD 将向 Meta 投资 10 亿美元，Helios 机架已安装测试
🧠Meta AI 总监离职创办 Ndoc，想把文档、数据与演示放进同一工作区
🏢金山办公 Q3 归母净利降 72.3%，经调整净利增 8.44%
💡微软 CEO 纳德拉：AI 更强，并不会自动变得更安全
🔋2999 元起，REDMI K100 发布：10000mAh 电池配双芯
🪄REDMI Book 2027 换新模具，14 英寸轻至 1.28kg
💬Rokid 新眼镜接入微信，手势戒指可盲操回复与翻页
⏰Google 发布世界模型 Genie 4：支持原生音频，模拟时长延长至两分钟
🧠微信输入法新增「AI 录音笔」，录音可实时整理成结构化网页
🔓百度开源 Unlimited OCR 2，称可一次解析 3600 页《哈利·波特》
🌱灵光 2.0 全面开放，可持续盯价、记住前情并跟进事务
💾英伟达 RTX 路径追踪工具更新，新增 Windows on Arm 支持
🪄富士 mini Evo 2 加入触屏与短片，11 月 27 日起发售
📷影石 Luna Ultra 获 Wirecutter 便携 Vlog 相机推荐
✨魏牌二代蓝山首秀，改用方盒子造型并提供五座、六座版本
🎮《刺客信条：黑旗 记忆重置》迎来最后一次大型更新，加入五轮新游戏＋
🔓蚂蚁百灵开源 Ling-3.1 增强版，支持 100 万 Token 上下文
🧠Liquid AI 发布 LFM3，小模型覆盖文本、图像、语音与决策
🎬Omdia：英国微短剧观众中，男性占 54%
🛍️泡泡玛特第三季度收益同比降 45%—50%，海外市场降幅更大
🎨红果已处置 20426 部低质同质化短剧，重点治理「换皮」与撞脸
👀HBO 接手开发剧版《群鸟》，《继承之战》莎拉·斯努克主演
苹果发布 M6 MacBook Pro 和 iMac
苹果发布搭载 M6 芯片的 MacBook Pro 和 iMac，国行起售价 分别为 15999 元、12499 元。
M6 配备 12 核 CPU 和 12 核 GPU，相比 M5 各增加 2 核，GPU 每个核心均带有神经网络加速器，并配备双 16 核神经网络引擎。苹果称，相比 M5，图形性能最高提升 30%，端侧 AI 工作负载性能最高提升 25%。
MacBook Pro 标配 16GB 统一内存，最高可选 32GB，内存带宽为 170GB/s。
iMac 外观则延续 24 英寸一体机设计，提供蓝、绿、粉、黄、紫、橙、银七种配色，配备 4.5K 视网膜显示屏；最高配备 4 个雷雳 4 端口，并支持最多 2 台 6K 外接显示器。
iPhone 18 Pro 上市三周在华销量增 29%，美国与日本下滑
Counterpoint Research 10 月 27 日公布的周度销量追踪显示，iPhone 18 Pro 系列上市前三周的全球销量较上一代同期增长 3.1%，主要增量来自中国市场：在华销量同比增长 29%，美国、日本市场则均出现下滑。
该机构认为，中国市场受益于上市初期供货改善，以及安卓竞品旗舰机型涨价。美国消费者对新机功能、设计升级的感知较弱；日本市场则受到汇率放大涨价幅度的影响。
iPhone 18 Pro Max 占本代 Pro 系列销量的 56%，略低于上一代的 57%。Counterpoint 副总监林秀晶表示，中国市场的强势是否来自提前集中购机，还需后续数周观察；iPhone Duo 能带来新增需求，还是分流 Pro Max，也是短期变量。
有线电视与 IPTV 新规：不得播放开机广告，首页须设免费专区
国家广播电视总局 10 月 27 日公布《有线广播电视运营服务管理规定》和《交互式网络电视运营服务管理暂行规定》，两份规章均将于 11 月 27 日施行。
新规要求相关运营服务提供者坚持直播优先，实现开机看全屏直播、设置频道回看入口，不得播放开机广告。用户应能通过便捷入口，自行选择开机进入直播或电视交互主页；智能电视机也须支持优先呈现直播频道。
在收费和操作方面，有线电视与 IPTV 首页须设立免费专区，重点推荐优质免费内容，精简收费栏目并明示收费信息；遥控器应能一键切换直播信源，终端还须持续缩短开机时长。
IDC：中国手机 Q3 出货下滑 15.2%，折叠屏连续四季度下降
IDC 10 月 27 日发布的季度跟踪报告显示，今年第三季度中国智能手机市场出货量约 5806 万台，同比下降 15.2%，降幅较上半年约 4% 进一步扩大。
IDC 将市场概括为「量跌价升」：存储及其他元器件成本上涨传导至终端，厂商调整配置、提高定价，消费者则延长换机周期。华为市场份额升至 21.1%，出货量保持同比两位数增长；荣耀是头部主流厂商中唯一实现出货量环比正增长的品牌。
折叠屏的调整更明显，第三季度中国市场出货约 177 万台，同比下降 32.7%，已连续四个季度下滑。华为在该市场的份额超过 70%，三星与小米分列其后。IDC 预计，随着苹果入局、更多品牌推出阔折叠新品，国内折叠屏市场将从下一季度起回暖。
张一鸣以 6700 亿元登顶胡润百富榜，梁文锋首次上榜排第七
胡润研究院 10 月 27 日发布《2026 胡润百富榜》，字节跳动创始人张一鸣以 6700 亿元财富第二次成为中国首富，较去年增加 2000 亿元，并刷新榜单的中国首富财富纪录。
农夫山泉创始人钟睒睒以 4100 亿元退居第二。DeepSeek 创始人梁文锋首次上榜，以 2800 亿元位列第七；宇树科技创始人王兴兴也首次上榜，财富为 695 亿元。
本次榜单共有 1529 位个人财富 50 亿元以上的企业家，比去年增加 95 位，总财富为 32 万亿元。胡润称，半导体、人工智能、机器人等行业是今年财富增长较突出的领域。榜单的财富计算截止日为 9 月 1 日，采用公开股权、市场价值及调研信息进行估算。
邓煜：AI 公司应公开证明思路，让用户决定数据是否用于训练
数学家邓煜在 10 月 25 日由《知识分子》《赛先生》联合举办的论坛演讲中谈到，AI 对数学研究有很大促进作用，但提供模型的公司同样需要面对学术规范、思想来源和用户权益问题。
他认为，如果把 AI 当作发表研究成果的研究者，就应解释重大证明是如何得到的，说明其中沿用了哪些人的思路。数学家报告成果时需要回答这些问题，AI 公司也应公开相应思维链，不能仅以防止蒸馏为由拒绝。
从商业关系看，邓煜质疑：数学家付费使用工具、输入研究思路后，模型提供方再利用这些思路训练产品并发布成果，用户的贡献应如何得到承认？他主张，是否允许输入数据参与训练，应由用户自行选择。
他同时强调，数学家不应拒绝使用 AI，而应把它作为推进研究的工具，由人提出问题、把握方向。他提出的方案包括建立思路和贡献的追溯机制，让成果能够说明某个想法最初由谁提出。
研究：豆包引用抖音的权重升至 31.8%
新榜服务发布的 AI 信源研究显示，在其监测样本中，豆包对抖音内容的引用权重由今年 5 月的 10.79% 升至 8 月的 31.8%，接近三分之一。
报告还称，百度在 6 月首次将 B 站和好看视频纳入所监测的引用信源，两者权重合计达到 12%；到 8 月，B 站权重稳定在 8.4%，排在第二位。
新榜进一步分析视频标题与引用次数的关系，发现标题与问题的语义相关性、是否包含品牌和专业术语，比单纯增加标题长度或堆砌规格数字更有解释力。这些结果来自新榜智汇的监测与统计，反映所覆盖问题和内容样本中的引用分布。
大众在海外召回约 7.27 万辆电动车，部分车型充电上限暂降至 80%
据第一财经报道，德国联邦汽车运输管理局公布的三批召回备案，合计涉及约 7.27 万辆大众集团电动车，包括部分大众 ID.3、ID.4、ID.7、ID.Buzz 和 Cupra Born，原因是高压电池电芯存在安全隐患。
大众英国官网确认，使用特定供应商于 2022 至 2025 年生产的高压电池的部分 ID.3、ID.4、ID.7 需要检查；电量超过 80% 时，电池可能过热并有起火风险。厂商已通过远程方式为多数受影响车辆暂时设置 80% 充电上限，后续将通知车主检查，必要时更换电池模组。
大众英国未披露供应商名称。第一财经指出，美国另一起涉及 22524 辆 ID.4 的召回文件将供应商列为 SK On，但不能据此认定所有地区涉事电池均来自该企业。报道发布时，中国市场尚未公布此次 ID. 系列召回公告。
《宝可梦大集结》海外版将于明年 3 月停服，保留资料馆
《Pokémon UNITE》日本官网 10 月 27 日公告，Nintendo Switch 版和手机版将于明年 3 月 23 日日本时间 9 时、即北京时间 8 时结束服务。官方当日已停止宝石销售，以及 UNITE 会员的新订阅和自动续费；现有宝石与会员权益可在相应期限内继续使用。
官方计划于 11 月 12 日加入「资料馆」功能，让玩家在停服后仍能查看已获得的电子服、时装、成就和部分比赛记录。玩家需要在服务结束前下载相关资料，停服后将无法再下载缺失数据。
此前公布的 Pokémon UNITE Aeos Crown 2027 Tokyo 赛事仍将按计划举办。未使用的付费宝石将在停服后开放退款申请，具体流程由官方另行说明。本次公告针对海外版。
红果否认「年收入 400 亿元」传闻，称大部分收入用于内容投入
红果短剧 10 月 27 日发布声明，否认网络流传的「年化收入逼近 400 亿元」「今年营收达 420 亿元至 450 亿元」等说法，称相关收入数据不实。
声明指出，网传内容援引的「机构数据」「行业数据」「专家预计」均非官方信息，由此作出的比较、推算和延伸解读也没有事实依据。
红果同时表示，目前平台大部分收入持续用于内容创作者分账、版权保底和优质内容扶持。
小米武汉首个员工住宅开盘，297 套房限内部员工购买
据《21 世纪经济报道》报道，小米在武汉的首个员工住宅项目「光谷梦享家」已开盘，项目规划 4 栋住宅、共 297 套房源，备案均价为每平方米 13885 元，购房对象限定为小米体系内已签订劳动合同、正常缴纳社保的在职正式员工。
项目距离小米武汉科技园约 700 米，提供约 110 平方米、122—125 平方米的三房，以及约 143 平方米的四房，按毛坯标准交付。报道援引小米生态链总经理陈波的介绍称，从项目步行到科技园需 10 分钟左右。
武汉市住房和城市更新局公示的预售资料显示，项目于 10 月 16 日取得预售许可，开发主体为小米全资子公司谧空间武汉信息科技有限公司，计划于 2028 年 5 月交付。
AMD 将向 Meta 投资 10 亿美元，Helios 机架已安装测试
AMD 宣布将参与 Meta 的股权发行，投资 10 亿美元。双方同时披露算力合作进展：首批 Helios 机架级系统已在 Meta 安装并进行测试，后续计划扩大部署。
AMD 的 EPYC 处理器为 Meta 的 Muse 个人 AI 智能体提供专用计算环境。Meta 已部署数百万颗不同代际的 EPYC，并成为第六代 EPYC 的首批重要客户；双方还在合作开发后续平台。
在 GPU 端，Meta 已将 Instinct 用于训练、排序与推荐系统，并计划部署针对自身工作负载优化的定制 GPU。软件合作覆盖 PyTorch、Triton 和 ROCm；Muse Spark Open 已获得 Instinct 首日支持，Muse Glimmer 则支持 Ryzen AI PC 与 Radeon GPU。
Meta AI 总监离职创办 Ndoc，想把文档、数据与演示放进同一工作区
据 TechCrunch 对罗希特·帕特尔（Rohit Patel）的采访，这位在 Meta 工作约 8 年的超级智能实验室前总监已离职，创办 AI 原生办公软件 Ndoc，希望把 Excel、Word、PowerPoint 等工具间分散的工作整合到同一界面。
Ndoc 将工作区分为三部分：Stage 展示幻灯片、仪表盘和迷你应用；Narrative 放分析、编号引用和图表；Model 保存底层数据集与 Python 脚本。用户上传数据后，可让 AI 分析并生成说明，也能检查计算脚本，或请 AI 解释公式所在的代码。
帕特尔称，文档可导出为浏览器能打开的单个「.ndoc.html」文件，他计划开放这一格式。团队目前有 3 人、尚未融资，正在开展企业试点，并向候补名单中的少量客户开放 alpha 版；注册免费，使用与 Token 消耗收费。beta 版计划在未来数周推出，完整产品计划明年发布。
金山办公 Q3 归母净利降 72.3%，经调整净利增 8.44%
金山办公 10 月 27 日披露三季报，第三季度营业收入为 17.66 亿元，同比增长 16.11%；归母净利润为 1.19 亿元，同比下降 72.3%，主要受对外投资基金项目投资收益降低影响。扣非净利润为 1.14 亿元，同比下降 72.85%。
公司披露的经调整归母净利润为 5.39 亿元，同比增长 8.44%。这一口径剔除了股份支付费用、基金相关投资收益，以及这些调整事项对应的所得税影响。
分业务来看：
- WPS 个人业务收入 10.50 亿元，同比增长 16.78%。公司称，全新版「灵犀」等 AI 产品带动了付费转化率及客单价提升。
- WPS 365 业务收入 2.70 亿元，同比增长 33.83%，企业级产品 WPS Comate 可连接企业数据与业务系统、执行任务，并管控数据和权限边界。
- WPS 软件业务收入 4.14 亿元，同比增长 5.78%。
第三季度经营活动现金流量净额为 4.67 亿元，同比下降 13.23%。截至 9 月底，WPS Office 全球月度活跃设备数为 6.78 亿，其中 PC 版为 3.41 亿，移动版为 3.37 亿。
💡 微软 CEO 纳德拉：AI 更强，并不会自动变得更安全
微软 CEO 萨提亚·纳德拉在 20VC 播客中说，他改变了一个过去相信的判断：模型能力提升后，AI 与人类目标的对齐会自然出现。这里的「对齐」，指让 AI 的行为符合人类的价值与目标。
据《商业内幕》对访谈的报道，主持人 Harry Stebbings 问他，过去一年改变了什么看法。纳德拉回答，模型扩展和能力进步大体符合他的预期，但对齐仍是一项独立的挑战。
我过去总以为，对齐会自然而然地出现。我现在不再相信这一点。我认为，对齐可能是一个更难、也不同的问题。
纳德拉此前于 10 月 10 日提出，企业应把 AI 模型当作潜在的内部威胁，并假设模型可能已被攻陷。对企业部署而言，他给出的办法是在非确定性模型外建立确定性的系统设计、人类控制和可靠的操作流程；现有行业标准不足时，还需要制定新标准。
2999 元起，REDMI K100 发布：10000mAh 电池配双芯
REDMI K100 于 10 月 27 日发布并开启预售，标准版首次搭载游戏独显 D2，配合高通骁龙 8 至尊版，主打游戏性能和大电池续航。
手机配备 6.83 英寸直屏，支持 165Hz 刷新率、全亮度 DC 调光及 3500nits 峰值亮度。官方称，屏幕已原生适配超过 40 款高帧率游戏，独显芯片可用于超分辨率与超帧处理。
续航方面，K100 内置 10000mAh 电池，支持 100W 有线快充及 27W 有线反充。小米实验室给出的 DOU 续航测试结果为 2.43 天。其他配置包括 5000 万像素主摄、3D 超声波指纹，以及出厂搭载的小米澎湃 OS 4。
12GB + 256GB、12GB + 512GB、16GB + 512GB 三个版本的双 11 限时特惠价分别为 2999 元、3499 元和 3899 元；首档叠加国补后的价格为 2549.15 元。新机将于 10 月 31 日 10 时正式开售，提供黑色、雪岩白和流萤绿三种配色。
REDMI Book 2027 换新模具，14 英寸轻至 1.28kg
REDMI 在 10 月 27 日的发布会上推出 Book 2027 系列，这是该数字系列时隔三年的模具更新。新品有 14 英寸和 16 英寸两种尺寸，提供月光银、深空灰、星云紫和海风蓝四种配色；14 英寸机型轻至 1.28kg，机身最厚处为 13.95mm。
系列搭载第三代英特尔酷睿处理器、LPDDR5X 7467MT/s 内存，采用双风扇散热，官方称整机稳定性能释放最高 30W。内置 72Wh 电池，官方模拟办公续航为 20.2 小时，标配 68W USB-C 电源适配器。
软件预装超级小爱 2.0，可调用小米 MiMo 模型完成 PPT、网页数据看板等任务；小米笔记 PC 版支持与手机、平板同步，也支持米家远程开关机和平板远程控制。新品已开启预约，11 月 3 日发售，14 英寸酷睿 5 320、16GB + 512GB 版本在适用地区的国补到手价为 5779.15 元。
Rokid 新眼镜接入微信，手势戒指可盲操回复与翻页
在 10 月 27 日的 Rokid Jungle 2026 发布会上，乐奇展示了新一代 AI 眼镜与配套交互：Rokid Glasses 2、双目全彩眼镜 Rokid Color，以及智能戒指「乐奇魔戒」。
乐奇已将微信接入眼镜端。收到消息后，通知可显示在镜片视野中，用户可通过语音，或搭配戒指的指尖手势确认并回复消息。
「乐奇魔戒」支持轻触、双击、滑动和长按等操作，可用于提词器翻页、切歌、抓拍以及确认 AI 指令，减少触摸镜腿或在公共场所开口操作的需要。
Google 发布世界模型 Genie 4：支持原生音频，模拟时长延长至两分钟
Google DeepMind 昨日发布新一代世界模型 Genie 4，支持实时生成画面、原生音频和物理模拟，单次交互模拟时长约为两分钟，较上一代的约一分钟有所延长。
用户输入一段文字即可生成可探索的虚拟世界，并通过键盘控制角色移动、调整镜头。在运行过程中，用户还能继续输入提示词，改变天气、添加人物或物体。
Genie 4 首次支持原生音频生成，可实时生成环境声音、音乐和人物对话，虚拟角色能根据用户输入和场景变化动态生成语言与反应。
物理模拟方面，谷歌展示了西瓜受冲击、枫树种子飘落、拥挤的乡村火车站等案例，涉及碰撞、空气动力学、重力和复杂人群行为。
🔗 相关阅读：刚刚，谷歌 Genie 4 发布，120 秒的世界却让我看见 AI 的未来
微信输入法新增「AI 录音笔」，录音可实时整理成结构化网页
微信派介绍，微信输入法新增「AI 录音笔」能力，可将录音实时转写为结构化网页，并按上下文整理重点和逻辑。
功能除了转成文字，还能修正错别字、删去冗余口头表达，并根据内容生成图表、图示和交互原型。微信派展示了会议纪要、直播内容回顾、家庭事务和旅行讨论等整理场景，总结页面可通过右上角的「分享」按钮发给好友。
手机端可从键盘工具栏或微信输入法 App 进入；电脑端可通过快捷键、Mac 顶部菜单栏或 Windows 工具栏打开。该功能支持多端同步，关联设备之间可以同步历史录音总结，用户可在最新版微信输入法中体验。
百度开源 Unlimited OCR 2，称可一次解析 3600 页《哈利·波特》
百度 10 月 27 日发布并开源文心 Unlimited OCR 2，主打超长文档的连续解析。百度称，模型可在一次任务中处理《哈利·波特》全系列 7 部作品、合计超过 3600 页，在文档增长后保持较稳定的解析精度与处理效率。
其核心是稀疏参考滑动窗口注意力机制 SR-SWA：先由页面门控缩小相关页面范围，再选取关键视觉 Token，减少对整份文档重复计算。配合无限解码机制，模型将单步解码所需的计算和显存开销控制在较稳定的范围内。
百度公布的 MPDocBench 和 MarathonDocBench 综合得分分别为 84.6 和 77.1。代码和模型权重已在 GitHub、Hugging Face 开放，开发者可使用 Transformers 或 SGLang 推理；百度智能云和 AI 开放平台也已上线相关能力。
灵光 2.0 全面开放，可持续盯价、记住前情并跟进事务
蚂蚁推出灵光 2.0，将其定位为「能管事的 AI 助理」，从单轮对话扩展到持续跟进的个人 Agent。用户关掉 App 后，任务仍可在独立云主机及其浏览器中运行，并在有结果或需要决策时提醒用户。
APPSO 的体验中，灵光可读取宠物保险合同、整理报销条件并设置后续提醒，也能记住此前沟通的情况；购物时可比较同款商品在不同平台的价格、折算优惠，并持续跟踪价格变化。旅行人数或天气变化后，它还能基于已有偏好调整路线和安排。
灵光 2.0 支持用户授权接入支付宝账单等服务，整理周期性扣款和消费记录，也可连接钉钉、飞书等工具。在这些体验场景中，信息整理、提醒和后续跟进由 AI 承担，最终决定仍交给用户。
🔗 相关阅读：中国版 Muse 能管事！我把宠物保单、双 11 算价和周末旅行全交给了灵光 2.0
英伟达 RTX 路径追踪工具更新，新增 Windows on Arm 支持
英伟达 RTX Path Tracing 工具已更新至 1.9.0。官方发布记录显示，新版相较 1.8.1 在画质相当或更好的条件下，整体性能约提升 20%，并新增 Windows on ARM64 支持。
本次优化包括将部分路径追踪任务与 DLSS 光线重建、后处理放在异步计算队列中重叠执行，在实例较多的场景降低 CPU 帧耗时，并减少加速结构的内存占用。
新版还默认启用 SHaRC 辐射缓存，以相近成本提供更多间接光照；NEE-AT 可处理的光源数量由 51.2 万提升至 200 万。该工具面向开发者提供实时路径追踪的示例和参考实现，1.9.0 的官方版本记录发布于 10 月 23 日。
富士 mini Evo 2 加入触屏与短片，11 月 27 日起发售
富士发布混合型拍立得相机 instax mini Evo 2，保留数码拍摄后挑选照片打印的方式，首次为 instax 系列加入触摸屏，4.5 英寸屏幕可用于调整设置及滑动浏览照片。
根据富士日本总部 10 月 27 日公告，新机最高照片分辨率提升至 1300 万像素，使用这一分辨率需插入 microSD 卡，并关闭「Expanded View」扩展取景模式。扩展取景可在屏幕中显示打印画框之外的区域，帮助判断移动主体何时进入构图。
新增「Immersive Movie」功能用于记录按下快门前的拍摄瞬间，生成 5 秒、9:16 竖屏短片。短片可在相机上回看，也可通过专用 App 传输至手机；相机仍保留 10 种镜头效果和 10 种胶片效果的组合。
mini Evo 2 将于 11 月 27 日起发售。富士北美公布的美国建议零售价为 289.95 美元。
影石 Luna Ultra 获 Wirecutter 便携 Vlog 相机推荐
《纽约时报》旗下 Wirecutter 在更新的《最佳 Vlog 相机及设备》指南中，将影石 Luna Ultra 列为便携、可随身携带的相机推荐，并建议希望获得一体化拍摄方案的创作者选择其创作者套装。
这份指南于 10 月 14 日更新。Wirecutter 看重的组合包括口袋机身内的云台防抖、主体追踪，以及可拆卸的 2 英寸 OLED 触屏。屏幕取下后可充当无线监看与遥控器，方便高举相机或从特殊机位拍摄；套装麦克风则在嘈杂场景下接受了人声收音测试。
指南提到，Luna Ultra 支持最高 8K 30 帧及 4K 120 帧录制，并提供 10-bit I-Log。其单机续航实测约 2 小时 45 分钟，搭配电池手柄可再延长约 1 小时。
Wirecutter 也指出，主摄视角未必适合大合影，数字变焦越深入，画质损失越明显；需要更宽视野或更多镜头选择的拍摄者，可考虑其他相机类型。
魏牌二代蓝山首秀，改用方盒子造型并提供五座、六座版本
魏牌二代蓝山于 10 月 27 日亮相，新车采用方盒子造型，提供五座和六座版本，动力改为 2.0T 插电式混合动力系统。
据《汽车行业关注》报道，新车使用直瀑式格栅、复古圆形大灯及外挂备胎，尾门采用侧开设计，车顶配有激光雷达。五座版车长为 5140mm，六座版加长至 5300mm，宽度、高度分别为 2050mm 和 1960mm，两种版本轴距均为 3010mm。
报道援引申报信息称，发动机最大功率为 182kW，前、后电机功率分别为 110kW 和 220kW。两种电池容量分别为 53.6kWh 和 64.7kWh，对应申报纯电续航 220 公里和 270 公里。
《刺客信条：黑旗 记忆重置》迎来最后一次大型更新，加入五轮新游戏＋
育碧宣布，《刺客信条：黑旗 记忆重置》最后一次大型内容更新将于北京时间 10 月 27 日 22 时推送，面向全平台加入「新游戏＋」、噩梦难度及新收集内容，后续数月还会继续提供体验优化更新。
新游戏＋可继承爱德华的工具、升级、已解锁武器，以及寒鸦号升级、已解锁船员技能和宠物等内容。玩家需要重新招募船员、夺取要塞，地图迷雾也会重置。每轮重玩会增加精英士兵和高等级巡逻船，最多可开启五轮新游戏＋，并加入专属货币兑换新装扮。
噩梦模式会提高敌人伤害、缩短格挡时机，并关闭防止单次致命攻击的保护机制；潜行中敌人发现玩家更快，水下生物伤害和氧气消耗也会增加。
更新还恢复了原版部分处决动作，加入隐藏社区宝箱、新服装及宠物外观。
蚂蚁百灵开源 Ling-3.1 增强版，支持 100 万 Token 上下文
蚂蚁百灵开源 Ling-3.1-flash-1019 模型权重，采用 MIT 许可证，支持 100 万 Token 上下文。模型总参数为 5600 亿，每个 Token 激活 250 亿参数，采用混合线性 MoE 架构。
官方模型卡称，新版通过进一步后训练加强推理与智能体任务能力，Artificial Analysis Intelligence Index v4.3.2 得分由前代的 41 提升至 44，AutomationBench-AA 得分由 62% 提升至 68%；这些是模型卡列出的评测结果。
部署文档提供 SGLang 与 vLLM 路径，支持思考开关、工具调用及 MTP 推测解码。官方推荐的 SGLang 示例使用 8 张 H200，并列出 4 张 GB300 的配置方式，主要面向服务端部署。
Liquid AI 发布 LFM3，小模型覆盖文本、图像、语音与决策
Liquid AI 发布 LFM3 系列，首批 6 款模型均处于约 20 亿参数规模，覆盖文本、视觉、音频、全模态及决策任务，共用 LFM3-1.7B 基础骨干，权重已在 Hugging Face 上线。
文本模型 LFM3-1.7B 支持 256K Token 上下文，4-bit 精度下约占 1GB 内存，可在手机和笔记本上本地运行，也能通过单张 GPU 微调。新架构结合门控短卷积、滑动窗口注意力和全局注意力，预训练使用了 22 万亿 Token；模型面向本地智能体、工具调用和特定任务定制。
官方模型卡同时给出适用边界：受规模限制，不推荐将文本模型用于复杂多步推理或智能体编程。系列采用自定义的 LFM Open License v1.0，商业使用受营收门槛限制。
Omdia：英国微短剧观众中，男性占 54%
Omdia 最新消费者研究显示，男性占英国微短剧用户的 54%。在受调查市场的微短剧用户中，男性占比为 47%，美国、西班牙、巴西分别为 50%、47% 和 29%，地区差异明显。
这项调查于今年 4 月开展。Omdia 媒体与娱乐业务主管 María Rúa Aguete 表示，行业长期把微短剧与女性观众、爱情题材联系在一起，可能低估了它的受众范围。随着市场扩大，喜剧、惊悚、犯罪和动作题材也有
本文由 AI宝库自动抓取并翻译，版权归原作者所有；如需查看原文请点下方链接。
15999 morning newspapers, apple publishM6 MacBook Pro/Zhang's new rule of $670 billion: no commercial openings
publish M6 MacBook Pro and iMac
📉iPhone 18 Pro shows an increase of 29% in sales in three weeks, US and Japan Slide
New rule for cable TV and IPTV: no commercials should be broadcast, and free-of-charge areas on the front page
📱IDC: Chinese mobile phone Q3
Zhang Zhang is on the list of 67 billion dollars, and Liang Wenzheng is on the seventh list.
🧠: AI should publicly prove the idea and let the user decide whether the data will be used for training
Research: the weight of Doubao citing Douyin rose to 31.8%.
The public recalled about 7.27 million electric vehicles overseas, with some of the vehicle-type ceiling temporarily reduced to 80 per cent.
The overseas edition of "Big Dreams" will be suspended in March next year and will keep the library.
The red fruit denies the rumour of "income 40 billion yuan a year" and claims that much of the income is spent on content inputs.
The first employee's house in Mivuhan is open, 297 rooms are open for internal staff.
🍎AMD will invest $1 billion in Meta and the Helios rack has been tested
🧠Meta AI Director General left to create Ndoc to put files, data and demonstrations in the same workspace
Q3 Net gain to mother 72.3%, adjusted 8.44%
微Microsoft CEO Nadra: A stronger AI does not automatically become safer
From 2999, REDMI K100 publish 10000 mAH Batteries with double cores
🪄REDMI Book 2027 New mold, 14 inches light to 1.28kg
WeChat with new glasses.
⏰Google publish World Model Genie 4: Support proto-audio, simulation extended to two minutes
Add "AI tape pen" to the Z60Z input method, which can be put into a structured network in real time Page
🔓 Baidu Open source Unlimified OCR 2, stating that 3,600 pages of Harry Potter
光 2.0 Full open, sustainable price-seeking, forward-looking and follow-up
Update RTX path tracker to add Windows on Arm support
Mini Evo 2 added touch screens and short films to be sold on November 27th.
Luna Ultra recommended by Wirecutter with Vlog camera
Wei's 2nd-generation Blue Mountain premiere, changed to a box with five, six versions.
"The Assassin Letter: Black Flag Reset" is the last big update to add five new rounds to the game.
Ling-3.1 enhanced version, supporting 1 million Token context
🧠Liquid AI publish LFM3, small model over text, images, voice and decision-making
🎬Omdia: Men account for 54% of the viewers of the British mini-short opera.
The third quarter of the year saw a 45-50 per cent decline in the earnings of bubbly Matt.
We've got 20,426 low-quality homogeneization shorts, with a focus on skin change and face bumps.
HBO takes over the development of The Birds, The War of Succession, by Sarah Snook.
Apple publish M6 MacBook Pro and iMac
MacBook Pro and iMac, apple publish, carrying M6 chips, with start-up prices of $15999 and $12499 respectively.
M6 equipped with 12 nuclear CPUs and 12 nuclear GPUs, an increase of 2 nuclears each compared to M5, each core of which has a nerve network accelerator and a double 16 nuclear nervous network engine. Apple claims that graphic performance is up to 30 per cent higher than M5 and end-side AI job load is up to 25 per cent higher.
MacBook Pro set 16GB unified memory with a maximum option of 32GB and a memory bandwidth of 170GB/s.
iMac continues with a 24-inch one-in-one design, providing seven colours of blue, green, powder, yellow, purple, orange, silver, with a 4.5-k retina screen; a maximum of four lightning 4 ports and support for up to two 6-K external monitors.
iPhone 18 Pro Three-week increase in sales 29%, U.S. and Japan Slide
The weekly sales tracking published on October 27th shows that the global sales of the iPhone 18 Pro series increased by 3.1 per cent over the same period of the previous generation, with the main increases coming from the Chinese market: Sales increased by 29 per cent over the same period, while the United States and Japan markets declined.
The Agency believes that the Chinese market has benefited from improvements in supply in the early days of the market, as well as from an increase in the price of the Android-style flagship aircraft. Consumers in the United States are less aware of the functionality of the new machine and of the upgrading of the design; the Japanese market is affected by increased exchange rate price increases.
iPhone 18 Pro Max accounts for 56 per cent of the current Pro series, slightly less than 57 per cent of the previous generation. Under-Secretary-General Counterpoint Lin Su-kyung said that the strength of the Chinese market came from early central purchasing machines, which required a few weeks of follow-up, and that iPhone Duo could generate additional demand or divert Pro Max, a short-term variable.
Cable TV and IPTV new rule: no commercials should be broadcast, and free of charge on the front page
On 27 October, the National Radio and Television Authority published the Regulations for the Management of Cable Radio and Television Operating Services and the Provisional Regulations for the Management of Interactive Network Television Operating Services, both of which will be implemented on 27 November.
The new regulations require the relevant operational service providers to adhere to the priority of live broadcasting, so that they can open the whole screen and set up a channel to look back at the portal and not broadcast the commercial. Users should be able to access the live or interactive television home page at their own choice through easy access; smart televisions should also support the priority presentation of live channels.
In terms of fees and operations, cable television and IPTV are required to set up free-of-charge areas on the front page, with a focus on recommending high-quality free content, streamlining the charge line and specifying the information on the charge; the remote control should be able to switch the live feed source by one key and the terminal must continuously shorten the opening time.
IDC: Chinese mobile, Q3.
The quarterly tracking report of publish by IDC on October 27 shows that China ' s smartphone market volume in the third quarter of this year was approximately 5,80.6 million, a decrease of 15.2 per cent over the same period and a further increase of about 4 per cent over the first half of the year.
The IDC summarizes the market as "a drop in price": an increase in the cost of storage and other meta-ware transfers to the terminal, a reconfiguration of the manufacturer, higher pricing, and an extension of the replacement cycle for consumers. The market share of China rose to 21.1 per cent, and the volume of exports remained equal to the double-digits; glory was the only brand among mainstream head producers that achieved a positive increase in the volume of exports.
The adjustment of the folding screen was even more pronounced, with the Chinese market producing about 1.77 million units in the third quarter, a decline of 32.7 per cent over the same period and a decline for the fourth consecutive quarter. China shares more than 70 per cent of the market, followed by Samsung and millet. IDC expects that the domestic folding-screen market will warm up from the next quarter as apples enter the market and more brands are launched.
Zhang's first time on the list is 7th.
On October 27th, publish, the founder of the byte jump, Zhang Zhang Zhang, made China ' s first wealth by 67 billion yuan, an increase of 200 billion yuan over last year, and updated China ' s top wealth record.
The founder of the Farmer Mountain Springs, Zhong Zhong Zhong Zheng, was second in terms of $41 billion. The founder of DeepSeek, Liang Wenfeng, was first on the list of 28 billion dollars for the seventh time, and the founder of Woo Tree Technology, Wang Hing Xing, was first on the list of 69.5 billion yuan in wealth.
This list has 1529 individual wealth entrepreneurs over $5 billion, an increase of 95 places over last year, and a total of $32 trillion. According to Hu Yun, the semiconductor, artificial intelligence, robotics, etc. are areas where wealth growth has been more pronounced this year. The date for wealth calculations on the list is 1 September and is estimated using open equity, market value and research information.
Dumbledore: AI should publicly prove the idea and let users decide whether the data is for training
Mathematician Deng Deng, in a forum speech on October 25, co-organized by Intellectuals Mr. Sai, said that AI was a major contributor to mathematical research, but that companies that provided models also needed to face academic norms, intellectual sources and user interests.
In his view, if AI were to be treated as a researcher to publish the results of the study, it would be useful to explain how the material evidence was obtained and to indicate who had followed it. Mathematicians need to answer these questions when reporting on results, and AI should also disclose the corresponding thinking chain and not refuse it simply to prevent distillation.
From a business relationship point of view, Deng questions: how should the contribution of the user be recognized when the model provider uses paid, using tools, input into research ideas, to train the product and publish results? He argued that it was up to the users to choose whether to allow input data to participate in training.
At the same time, he stressed that mathematicians should not refuse to use AI, but rather use it as a tool to advance research, with people asking questions and taking direction. His proposals included the establishment of a retroactive mechanism of thinking and contribution, so that the outcome could indicate who had initiated an idea.
Research: the weight of Doubao citing Douyin rose to 31.8 per cent.
In its monitoring of sample, the AI source study of publish shows that the weight of Doubao ' s reference to Douyin content rose from 10.79 per cent in May to 31.8 per cent in August, almost one third.
The report also states that Baidu included station B and good video for the first time in June in the monitored citation source, with a combined weight of 12 per cent; by August, B station weight had stabilized at 8.4 per cent, ranking second.
It further analyses the relationship between the video title and the number of references, and finds that the semantic relevance of the title to the issue and the inclusion of brand names and professional terms are more explanatory than simply adding the length of the title or the number of stacking specifications. These results are derived from the monitoring and statistics of the new resource, reflecting the issues covered and the Quoting distribution of the content sample.
The public recalled about 7.27 million electric vehicles abroad, with some of the vehicle-type caps temporarily reduced to 80 per cent.
According to the first financial report, three sets of recalls issued by the German Federal Motor Transport Administration, involving a total of approximately 7.27 million mass-group electric vehicles, including some members of the public, ID.3, ID.4, ID.7, ID.Buzz and Cupra Born, were filed because of security risks to high-voltage battery cores.
Mass British official networks have confirmed that the parts of high-pressure batteries produced by specific suppliers in 2022 to 2025 id.3, ID.4, ID.7 need to be inspected; batteries may overheat and have a fire risk when more than 80 per cent of electricity is in excess. The manufacturer has temporarily set an 80 per cent charge limit for most of the vehicles affected by remote means, after which it will notify the owner of the vehicle and replace the battery module if necessary.
The name of the supplier was not disclosed by the public in the United Kingdom. The first financial note states that another recall in the United States involving 22524 ID.4 listed the supplier as SK On, but that it was not possible to conclude that all the area-related batteries originated from the enterprise. At the time of reporting on publish, the Chinese market had yet to publish the recall announcement of the ID. series.
The overseas edition of "Big Dream" will be suspended in March next year and will keep the library.
Pokémon UNITE announcement of October 27th, the Nintendo Switch and mobile phone versions will end next March 23rd at 9 p.m., Beijing time 8. The sale of gemstones was stopped on the official day, as well as new subscriptions and automatic renewals for UNITE members; existing gemstones and members ' rights may continue to be used for the corresponding period.
On November 12th, the official plan is to introduce the "Turkbook" function, so that players can still view the acquired electronic clothes, fashion, achievements and some of the records of the competition after they have stopped. Players need to download the relevant information before the service closes and will not be able to download the missing data after they have stopped.
The previously published Pokémon UNITE Aeos Crown 2027 Tokyo will continue as planned. The unused paid gemstone will open the application for refunds after the suspension, the process of which will be described separately by the official. This bulletin is aimed at overseas editions.
The red fruit denies the rumour of "income 40 billion yuan a year" and claims that most of the income is spent on content inputs.
On October 27, publish stated that it denied that the Internet's "annual income is approaching $40 billion" and that this year's revenues ranged from $42 billion to $45 billion, claiming that the relevant income data were false.
According to the statement, the reference to "institutional data" "industry data" "experts expect" is not official, and there is no factual basis for the comparison, extrapolation or extension.
At the same time, it is indicated that most of the current revenue of the Platform is used for content creators ' splits, copyright bottoms and quality content support.
Mivuhan's first employee's house is open, 297 flats are limited to internal staff.
According to the Economics Report 21, Mi's first employee housing project in Wuhan, "The Guangya Dream Home", has been opened, with 4 housing units, a total of 297 housing units, with an average filing price of $13885 per square metre, and a limited number of regular working employees who have entered into a labour contract within the millimetre system.
The project is approximately 700 metres from the Mivuhan Science and Technology Park, providing three rooms of approximately 110 square metres, 122-125 square metres, and four rooms of approximately 143 square metres, to be delivered at gross standards. According to the report, it took about 10 minutes to walk from the project to the Science and Technology Park.
According to advance sales information published by the City Housing and Urban Renewal Board of Wuhan, the project was granted a pre-sale licence on 16 October and was developed by the All-Smee Company, Space Wuhan Information Technology Ltd., scheduled for delivery in May 2028.
AMD will invest $1 billion in Meta and the Helios hangar has been tested
AMD announced that it would participate in Meta's equity distribution, investing $1 billion. The parties simultaneously disclosed the progress of the calculus cooperation: the first Helios shelf-level systems had been installed and tested in Meta, with plans to expand their deployment.
The AMD EPYC processor provides a dedicated computing environment for Muse AI AI agent of Meta. Meta has deployed millions of EPYCs across generations and has been the first major client of the sixth generation of EPYC; both parties are also cooperating in the development of a follow-up platform.
At the GPU end, Meta has used Instinct for training, sorting and recommending systems, and plans to deploy customized GPUs for its own load optimization. Software collaborations cover PyTorch, Triton and ROCm; Muse Spark Open has first-day Instinct support, and Muse Glimmer supports Ryzen AI PC and Radeon GPU.
Meta AI Director General left office to create Ndoc, trying to put files, data and demonstrations in the same workspace.
According to a interview with Rohit Patel by TechCrunch, the former director of the Super-Intelligence Laboratory, who had been working in Meta for about eight years, had left his post and created the AI Native Office software Ndoc, hoping to integrate the scattered work between Excel, Word, PowerPoint and others into the same interface.
Ndoc divides the work into three parts: Stace display slides, dashboards and mini-applications; Narrative Analytics, Numbering Quotes and Charts; Model saves bottom datasets and Python scripts. The user uploads the data to allow AI to analyse and generate instructions, to check the calculation script, or to ask AI to explain the code where the formula is located.
Patel claims that the document can be exported to a single ".ndoc.html" file that the browser can open, and he plans to open this format. The team is currently 3 people, not yet funded, is piloting the business and opening the alpha version to a small number of clients on the alternate list; registration is free of charge, using the Token consumption fee. The Beta version is to be launched in the coming weeks, and the whole product is to be publish next year.
Q3 QM 72.3 % + 8.44 %
On October 27, the Kim Shan office disclosed a three-quarter bulletin with a third quarter operating income of RMB 1,766 million, an increase of 16.11 per cent over the same period, and a net profit to the mother of RMB 119 million, a decrease of 72.3 per cent over the same period, mainly affected by a decrease in the return on investment in the OFDI fund project. The deduction of non-net profits was $114 million, a decrease of 72.85 per cent over the same period.
The company disclosed an adjusted net profit of $539 million, an increase of 8.44 per cent over the same period. This calibre excludes the shareholding of paid, the return on the related investment of the Fund and the effect of these adjustments on income tax.
For business purposes:
- WPS personal business income of RMB 1050 billion, an increase of 16.78 per cent over the same period. According to the company, new versions of AI products, such as Narcs, have led to paid conversion rates and to an increase in the unit price of customers.
- WPS 365 Business income of $270 million, an increase of 33.83 per cent over the same year, enterprise-level product WPS Company can connect business data to business systems, perform tasks and control data and permission borders.
- WPS software business income of $414 million, an increase of 5.78 per cent over the same period.
Net cash flows from operating activities in the third quarter were $467 million, a decrease of 13.23 per cent over the same period. By the end of September, the number of WPS Office monthly active devices worldwide was 678 million, of which the PC version was 341 million and the mobile version was 337 million.
💡 Microsoft CEO Nadra: A stronger AI does not automatically become safer
In his 20VC podcast, Microsoft CEO Satya Nadra said that he had changed a previously believed judgement that, with enhanced modelling capacity, the alignment of AI with human targets would come. Here's the "coherence" to make AI act in line with human values and goals.
According to the interview in The Business Inside, the moderator Harry Stebbings asked him what had changed in the past year. Nadra replied that the model expansion and progress in capacity largely met his expectations, but that alignment remained a stand-alone challenge.
I used to think that alignment would come naturally. I don't believe that anymore. I believe that alignment may be a more difficult and different issue.
Nadra had previously proposed on 10 October that enterprises should consider the AI model as a potential internal threat and assume that it might have been attacked. In the case of business deployments, he offered the option of establishing definitive systems design, human control and reliable operational processes outside the non-certainty model; new standards were needed when existing industry standards were inadequate.
From $2999, REDMI K100 publish:10000 mAH Battery with double core
REDMI K100 opened the pre-sale on October 27, publish, and the standard version first mounted the game uniquely D2, in conjunction with the Gotton Tyrone 8-to-Hork, with the main player playing game performance and large batteries.
Cell phones are equipped with 6.83 inches of direct screens to support 165Hz refresh rate, full brightness DC luminousity and 3500 nits peak brightness. Officially, the screen is adapted to more than 40 high frame games, and a unique chip can be used for ultra-resolution and hyperframe processing.
For the continuation of the voyage, K100 has a 10000 mAh battery to support 100W wired fast charge to 27W wired inverse. The results of the DOU renewal test given by the milli lab were 2.43 days. Other configurations include 50 million pixels, 3D ultrasound fingerprints, and milled milled milled millet OS 4.
12GB + 256 GB, 12GB + 512 GB, 16GB + 512 GB, 11 limited-time, 2999, 3499 and 3899 respectively, respectively, and 2549.15 dollars after the initial stacking. The new machine will be officially opened at 10 a.m. on October 31st, and will provide three colours: black, snow white, and fluent green.
REDMI Book 2027 for new mold, 14 inches light to 1.28kg
REDMI launched the Book 2027 series at the publish meeting on October 27th, a three-year update of the series. The new product is 14 inches and 16 inches, providing four colours of moon silver, deep ash, nebula and seawind; 14 inches light up to 1.28 kg, with 1395 mm at the height.
The series carries the third generation of Intel Cool Wisdom Processors, LPDDR5X 7467MT/s RAM, with a double fan heat spread, officially known as the whole machine stability release up to 30 W. An internal 72Wh battery with an official analogue office extension of 20.2 hours with 68W USB-C power adaptor.
The software pre-assembly Super Love 2.0, which enables the deployment of the MiMo model to PPT, web-page data boards, etc.; the Mi Note PC version supports synchronization with mobile phones, tablets, and also supports the remote control of the rice switch and board. New items have been opened for sale on November 3rd, and 14 inches of coolness 5,320, 16GB + 512GB copies have been paid at a hand price of $5,779.15 in the countries of the area.
Rokid's new glasses are connected to WeChat, and the sign ring can be used blindly.
At the Rokid Jungle 2026 publish meeting on October 27, Lucky presented a new generation of AI glasses interacting with each other: Rokid Glasses 2, double-eyed full-colour glasses Rokid Color, and the smart ring "Lage Ring".
Lucky's got WeChat in his glasses. Upon receipt of the information, the notification may show that, in the lens view, the user can confirm and respond to the information by voice or finger gestures with the ring.
The "Lage Ring" supports light touch, double-click, slide and long press, which can be used to use the terminograph to flip pages, cut songs, tap and confirm AI commands to reduce the need to touch mirror legs or open mouths in public places.
Google publish World Model Genie 4: Support proto-audio, simulation length extended to two minutes
Google DeepMind yesterday, publish's new-generation world model Genie 4, which supports real-time generation of images, proto-audio and physical simulations, with single interactive simulations lasting about two minutes, which is about one minute longer than the previous generation.
A user enters a text that generates an explorable virtual world and controls the movement of roles and adjusts shot through keyboards. In the course of the operation, the user can also continue to enter prompt, changing the weather, adding characters or objects.
Genie 4 supports primary audio generation for the first time, generating sound, music and person dialogue in real time, and virtual role generation of language and response based on user input and scene dynamics.
In terms of physical simulations, Google shows examples of watermelon shocks, seed drops of maple trees and crowded rural railway stations, involving collisions, aerodynamics, gravity and complex crowd behaviour.
Read about it: Just now, Google Genee 4 publish, 120 seconds of the world has made me see the future of AI
WeChat input method adds an "AI tape pen" which can be organized in real time into a structured network Page
According to WeChat, the WeChat input method adds an "AI tape pen" capability, which allows the recording to be translated into a structured web page in real time and to organize the focus and logic of context.
The function, in addition to converting to text, corrects miswords, deletes redundant oral presentations and generates charts, diagrams and interactive prototypes based on content. The WeChat team presented the wrap-up scenes of the proceedings, a live review, family affairs and travel discussions, and the wrap-up page was sent to friends via the "Share" button at the top right corner.
Cell phones can be accessed from the keyboard toolbar or the WeChat input method App; the computer can be opened through the shortcut, the Mac top menu bar or the Windows toolbar. This function supports multi-end synchronization, which allows for the synchronization of historical audio summaries between associated devices and users to experience in the latest version of the WeChat input method.
Open source Baidu Unlimited OCR 2, stating that 3,600 pages of Harry Potter
Baidu 27 October publish with open source Ernie Unlimided OCR 2, where the main type of continuous resolution of a very long document. According to Baidu, the model can handle the entire Harry Potter series of 7 works, with a total of more than 3,600 pages, in a single mission, and maintain a more stable resolution accuracy and processing efficiency following document growth.
The core is the thin-referenced slide window attention mechanism SR-SWA: close the relevant page range with the page door, then select key visual Token and reduce double counting of the entire document. Together with an unlimited decoder mechanism, the model controls the calculation and VRAM expenses required for one-step decodering within a more stable range.
The combined MPDocBech and MarathonDocBech scores were 84.6 and 77.1 respectively, published by Baidu. Codes and model weights are already open in GitHub, Hugging Face, and developers can use Transformers or SSGLang reasoning; Baidu smart clouds and AI open platforms are also online.
2.0 Full open, sustainable price-seeking, forward-looking and follow-up
Ants have launched luminous light 2.0 and positioned it as "AI Assistant in charge", extending from a single-wheeled conversation to an individual, Agent. After the user has turned App off, the task can still be run in the standalone cloud host and its browser, and alerts the user when there is a result or a decision is needed.
In the APPSO experience, it is also possible to read pet insurance contracts, organize reimbursement terms and set up follow-up reminders, keeping in mind previous communications; when shopping, it is possible to compare the prices of the same commodity with those of different platforms, to convert them and to keep track of price changes. It can also adjust routes and arrangements on the basis of pre-existing preferences as a result of changes in travel numbers or weather.
Psychic 2.0 Supports user-authorised access to services such as payment of valor bills, preparation of cyclical deductions and consumption records, and links to tools such as nails, flying books, etc. In these experience scenarios, the collation, alerting and follow-up of information is done by AI and the final decision remains with the user.
🔗 Read about: The Chinese version Muse can manage! I gave the pet policy, double 11 and weekend travel to the light. 2.0.
New Windows on Arm
The RTX Path Trading tool has been updated to 1.9.0. Official publish records show that the new version compares 1.8.1 with an overall performance increase of approximately 20 per cent under conditions of comparable or better quality and with new Windows on ARM64 support.
This optimization includes the re-establishment of partial routing tasks with DLSS light, the overlapping of reprocessing in an anisocalation line, the reduction of CPU frame time in more cases and the reduction of memory occupancy of accelerated structures.
The new version also defaults on the use of SHARC radiation cache to provide more indirect light at close cost; the number of NEE-AT-processable light sources increased from 51.2 million to 2 million. The tool is designed to provide developers with examples and references for real-time path tracking. The official version of 1.9.0 records publish on 23 October.
Fuji Mini Evo 2 joined the touchscreen and the short film, which was released from November 27th.
Instax mini Evo 2, the publish hybrid camera, maintains the digital method of selecting photo printing, adding a touch screen for the instax series for the first time, and 4.5 inches screens can be used to adjust settings and slide through photos.
According to the October 27 announcement by Fuji Japan headquarters, the top photo of the new machine, resolution, was raised to 13 million pixels, using this resolution to insert a microSD card and close the Expanded View extension. Extend the view to show the area outside the print frame on the screen to help judge when the moving subject enters the image.
Add "Immersive Movie" to record the moments before pressing the fast door, generating 5 seconds, 9:16 vertical clips. Short films can be viewed on the camera, and can be transmitted to a mobile phone via a special App; the camera retains a combination of 10 shot effects and 10 film effects.
Mini Evo 2 will be sold on November 27th. Fuji North America published a United States retail offer of $289.95.
Image Luna Ultra recommended by Wirecutter with Vlog camera
Under the New York Times, Wirecutter, in an updated guide to Best Vlog Cameras and Devices, recommended the stone Luna Ultra as a portable, portable camera and suggested that creators who wished to get an integrated filming programme choose their creator ' s suit.
This guide was updated on October 14. Wirecutter's heavyweight combination consists of a drowsy roof in the pocket, a main tracker, and a 2-inch OLED touch screen that can be removed. The screen is removed and used as a wireless viewer and remote control, which facilitates high-altitude cameras or shoots from special slots; the package of microphones is subject to human sound testing in a noisy scene.
The Guide mentions that Luna Ultra supports a maximum of 8K 30 frame and 4K 120 frame recording and provides 10-bit I-Log. Its single aircraft continues to be measured for approximately 2 hours and 45 minutes, with battery handles for an additional period of about 1 hour.
Wirecutter also pointed out that the main camera perspective may not be suitable for a large photo picture, and that the more the digital focal point becomes, the more the graphic loss becomes apparent; other camera types may be considered if there is a need for a wider view or more of the photographers chosen by shot.
Wei's 2nd-generation Blue Mountain premiere, modeled in a box with five, six versions.
On October 27, when Blue Mountain Wei was in the sun, the new car was modeled in a square box, providing five and six versions of power to a 2.0T plug-in hybrid power system.
According to Automotive Industry Concern, the new car uses a straight falls fence, a retrovert light and a spare tire, with a side-drive design for the tail door and laser radar on top. The five editions are 5140 mm and the six versions are increased to 5,300 mm, with a width of 2050 mm and a height of 1960 mm, respectively, with an axle distance of 3010 mm.
It was reported that the maximum power of the engine was 182 kW and the former and the post-electric power was 110kW and 220 kW respectively. The two batteries have a capacity of 53.6 and 64.7 kWh, respectively, corresponding to a declared total of 220 and 270 km.
Assassin Letter: Black Flag Memory Reset is the last big update to add five new rounds to the game.
The last large-scale update of the Assassin Letter: Black Flag Replacing will be delivered on October 27, 22nd, Beijing time, to add "New Game Plus" to the entire platform, the difficulty of nightmares and new collections, and will continue to provide an optimised update for months to come.
The new game + can inherit Edward's tools, upgrades, unlocked weapons, and the upgrading of the Grail, unlocked crew skills and pets. Players need to re-recruit crews and take forts, and map fogs can be replaced. Each round of heavy play will increase the number of elite soldiers and high-level patrol boats, with a maximum of five new rounds of games plus, and the addition of a new exclusive currency-exchange costume.
The nightmare model increases enemy injury, shortens block times, and closes protective mechanisms against single lethal attacks; the enemy finds players faster in the submersible, increases underwater biological harm and oxygen consumption.
The update also restored some of the executions in their original form, adding the hidden community treasure box, new clothing and pet look.
Ling-3.1 enhanced version, supporting 1 million Token context
Ling-31-flash-1019 model weights, MIT licence to support 1 million Token context. The total model parameter is 56 billion, each Token activates 25 billion and uses a hybrid linear MoE structure.
The official model card states that the new version enhances reasoning and AI agent mission capabilities through further post-training and that the Artificial Analysis Inteligence v. 4.3.2 scores rose from 41 in the previous generation to 44 and the Automation Bench-AAA scores from 62 per cent to 68 per cent; these are the results of the model cards.
Deployment documents provide SGLang and vLM paths to support thinking switches, tool call and MTP guess decoded. The officially recommended SGLang example uses 8 H200 and lists 4 GB300 configurations, mainly for service-end deployment.
Liquid AI publish LFM3, small model covering text, images, voice and decision-making
Liquid AI publish LFM3 series, the first 6 models are of approximately 2 billion parametric size, covering text, visual, audio, total mosaics and decision-making tasks, sharing LFM3-1.7B base cores, and weightings have been placed on Hugging Face.
The text model LFM3-1.7B supports 256K Token context, 4-bit with approximately 1GB memory, which can be operated locally on mobile phones and notebooks, and can be fine-tuned by single GPU. The new architecture combines door-control short volume, slide window attention and global attention, with 22 trillion token used for pre-training; the model is tailored to local AI agent, tool call and specific tasks.
The official model card also gives the applicable boundary: subject to size constraints, text models are not recommended for complex multistep reasoning or AI agent programming. A custom LFM Open License v. 1.0 is used in the series, and commercial use is subject to the threshold of collection.
Omdia: Men account for 54% of the viewers of the British microtrip.
The most recent consumer study in Omdia shows that men account for 54 per cent of the users of the United Kingdom's micro-short drama. The share of men among the users of the mini-short dramas in the surveyed market was 47 per cent, with significant regional differences of 50 per cent, 47 per cent and 29 per cent in the United States, Spain and Brazil, respectively.
This survey was launched in April this year. The Omdia Director of Media and Entertainment, María Rúa Aguete, stated that the industry ' s long-standing association with women ' s viewers and love stories may underestimate its audience. As the market expands, comedy, horror, crime, and action.
Auto-fetched and translated by AI Gem Box. All rights belong to the original author.