Company news and people to follow

Search opens wider

[Verified ^{{ˈverɪfaɪd 已核实的}} fact] SearchPro ^{{sɜːtʃ prəʊ 搜索接口名}} returns ranked ^{{ræŋkt 已排序的}} search results in JSON ^{{ˈdʒeɪsən 数据格式}}. Standard ^{{ˈstændəd 标准的}} and flagship ^{{ˈflæɡʃɪp 旗舰}} modes ^{{məʊdz 模式}} add specialist ^{{ˈspeʃəlɪst 垂类的}} and mixed ^{{mɪkst 混合的}} search; flagship ^{{ˈflæɡʃɪp 旗舰}} also adds authority ^{{ɔːˈθɒrəti 权威性}} scores ^{{skɔːz 评分}} and sublinks ^{{ˈsʌblɪŋks 子链接}}. The API ^{{ˌeɪ piː ˈaɪ 程序接口}} page was updated ^{{ʌpˈdeɪtɪd 已更新的}} on 20 October at 02:26:15, without a stated ^{{ˈsteɪtɪd 明示的}} time zone ^{{zəʊn 时区}}; its protocol ^{{ˈprəʊtəkɒl 协议}} version ^{{ˈvɜːʃən 版本}} is dated 8 May 2025. Earlier search services already existed. “Newly ^{{ˈnjuːli 新近地}} opened” therefore ^{{ˈðeəfɔː 因此}} needs the context ^{{ˈkɒntekst 背景}} of a product upgrade ^{{ˈʌpɡreɪd 升级}}, rather than a business built from nothing.

[Company claims ^{{kleɪmz 声称}}] Hundreds of millions of daily calls and more than 50 internal ^{{ɪnˈtɜːnəl 内部的}} businesses show use inside Tencent ^{{ˈtensent 腾讯}}, not outside paid demand ^{{dɪˈmɑːnd 需求}}. The ima and Tencent ^{{ˈtensent 腾讯}} Video cases report better source ^{{sɔːs 来源}} recall ^{{rɪˈkɔːl 召回率}} and information coverage ^{{ˈkʌvərɪdʒ 覆盖}}, but lack public samples and a controlled ^{{kənˈtrəʊld 设对照的}} outside test.

[Third-party report] Reuters ^{{ˈrɔɪtəz 路透社}}, in a 20 October report carried in full by Futu ^{{ˈfuːtuː 富途}} at 17:00, says ByteDance ^{{ˈbaɪtdɑːns 字节跳动}}, Alibaba ^{{ˌæliˈbɑːbə 阿里巴巴}} and Tencent ^{{ˈtensent 腾讯}} together ordered over 3 million RTX ^{{ˌɑː tiː ˈeks 显卡系列名}} PRO 5500 cards with a total sticker ^{{ˈstɪkə 标价标签}} value above US$35 billion ^{{ˈbɪljən 十亿}}. It relies ^{{rɪˈlaɪz 依靠}} on unnamed ^{{ˌʌnˈneɪmd 匿名的}} sources ^{{ˈsɔːsɪz 来源}}. Nvidia ^{{ɛnˈvɪdiə 英伟达}} could not promise ^{{ˈprɒmɪs 承诺}} full supply ^{{səˈplaɪ 供应}}; the report puts early samples in December and commercial ^{{kəˈmɜːʃəl 商业的}} delivery ^{{dɪˈlɪvəri 交付}} from Q1 2027. Tencent’s ^{{ˈtensents 腾讯的}} own share ^{{ʃeə 份额}}, deposits ^{{dɪˈpɒzɪts 订金}}, binding ^{{ˈbaɪndɪŋ 有约束力的}} terms ^{{tɜːmz 条款}} and cash ^{{kæʃ 现金}} payments ^{{ˈpeɪmənts 付款}} remain unknown ^{{ˌʌnˈnəʊn 未知的}}. These are three buyers ^{{ˈbaɪəz 买方}}’ reported ^{{rɪˈpɔːtɪd 报告的}} orders, not Tencent’s ^{{ˈtensents 腾讯的}} booked ^{{bʊkt 已入账的}} capital ^{{ˈkæpɪtl 资本}} spending.

[Cross-check / analysis ^{{əˈnæləsɪs 分析}}] Nvidia’s ^{{ɛnˈvɪdiəz 英伟达的}} own page names the Workstation ^{{ˈwɜːksteɪʃən 工作站}} Edition ^{{ɪˈdɪʃən 版本}} and says “Coming Soon”; this supports product identity ^{{aɪˈdentəti 身份}}, not allocation ^{{ˌæləˈkeɪʃən 分配}} to Tencent ^{{ˈtensent 腾讯}} or a specific ^{{spəˈsɪfɪk 具体的}} export ^{{ˈekspɔːt 出口}} licence ^{{ˈlaɪsəns 许可}}.

Temu reports over 13,300 removed ^{{rɪˈmuːvd 已移除的}} listings ^{{ˈlɪstɪŋz 商品链接}}

[Verified ^{{ˈverɪfaɪd 已核实的}} terms ^{{tɜːmz 条款}} / limits ^{{ˈlɪmɪts 限制}}] The 48-hour promise ^{{ˈprɒmɪs 承诺}} concerns a response ^{{rɪˈspɒns 响应}} to escalated ^{{ˈeskəleɪtɪd 升级处理的}} IP cases, not a decision, removal ^{{rɪˈmuːvəl 下架}} or final settlement ^{{ˈsetəlmənt 解决}}. IACC member brands ^{{brændz 品牌}} can use the special channel ^{{ˈtʃænəl 渠道}} without an extra ^{{ˈekstrə 额外的}} programme ^{{ˈprəʊɡræm 项目}} fee ^{{fiː 费用}}; ordinary ^{{ˈɔːdɪnəri 普通的}} Temu IP channels ^{{ˈtʃænəlz 渠道}} remain open to other rights holders ^{{ˈhəʊldəz 持有人}}. Neither the exact start of this response ^{{rɪˈspɒns 响应}} promise ^{{ˈprɒmɪs 承诺}} nor the cut-off date behind 13,300 was given. No total listing ^{{ˈlɪstɪŋ 商品链接}} count or repeat-listing ^{{rɪˌpiːt ˈlɪstɪŋ 重复上架}} figure was disclosed ^{{dɪsˈkləʊzd 披露的}}.

New products for 22–25 October

[Terms ^{{tɜːmz 条款}} / timing ^{{ˈtaɪmɪŋ 时间安排}}] Any purchase ^{{ˈpɜːtʃəs 购买}} qualifies ^{{ˈkwɒlɪfaɪz 符合条件}} for a LABUBU game number; spending RMB ^{{ˌɑːremˈbiː 人民币}} 69 qualifies ^{{ˈkwɒlɪfaɪz 符合条件}} for a separate ^{{ˈsepərət 单独的}} card offer. The in-store game has two daily sessions ^{{ˈseʃənz 场次}} of 20 places each, with limited gifts. GONG’s three roughly ^{{ˈrʌfli 大约}} 15 cm products use a different IP; their RMB ^{{ˌɑːremˈbiː 人民币}} 109 price must not be attached ^{{əˈtætʃt 附加的}} to Angry Molly. At the report cut-off, all selling or event dates were still in the future.

[Community ^{{kəˈmjuːnəti 社区}} comments ^{{ˈkɒments 评论}}] On the official ^{{əˈfɪʃəl 官方的}} Xiaohongshu note, “酒” (20 October 17:17) opposed a price rise; “香芋芝士xixi” (18:02) hoped for RMB ^{{ˌɑːremˈbiː 人民币}} 79; “Youyouyou” (19:20) asked for better quality ^{{ˈkwɒləti 质量}}, no price rise and a smaller size. Positive ^{{ˈpɒzətɪv 正面的}} buying interest was visible ^{{ˈvɪzəbəl 可见的}} too. These are ordinary ^{{ˈɔːdɪnəri 普通的}} users ^{{ˈjuːzəz 用户}} with unverified ^{{ˌʌnˈverɪfaɪd 未核实的}} backgrounds and self-selected ^{{ˌself sɪˈlektɪd 自行参与的}} comments ^{{ˈkɒments 评论}}, not paid-order data ^{{ˈdeɪtə 数据}} or a representative ^{{ˌreprɪˈzentətɪv 有代表性的}} survey ^{{ˈsɜːveɪ 调查}}. RMB ^{{ˌɑːremˈbiː 人民币}} 79 is a user’s ^{{ˈjuːzəz 用户的}} wish, not company pricing ^{{ˈpraɪsɪŋ 定价}}.

An intention ^{{ɪnˈtenʃən 意向}} is not a trade ^{{treɪd 交易}}

[Identity ^{{aɪˈdentəti 身份}}] The account ^{{əˈkaʊnt 账户}} is attributed ^{{əˈtrɪbjətɪd 归因于}} to Duan Yongping by independent ^{{ˌɪndɪˈpendənt 独立的}} public reporting ^{{rɪˈpɔːtɪŋ 报道}}.

A faster test needs a common yardstick ^{{ˈjɑːdstɪk 衡量标准}}

[Primary ^{{ˈpraɪməri 主要的}} evidence / correction ^{{kəˈrekʃən 纠正}}] The DeepSeek ^{{ˈdiːpsiːk 深度求索}}-V4.1-Flash model ^{{ˈmɒdl 模型}} configuration ^{{kənˌfɪɡəˈreɪʃən 配置}} lists 384 routed ^{{ˈruːtɪd 路由的}} experts ^{{ˈekspɜːts 专家}}, one shared ^{{ʃeəd 共享的}} expert ^{{ˈekspɜːt 专家}} and six active ^{{ˈæktɪv 激活的}} experts ^{{ˈekspɜːts 专家}} per token ^{{ˈtəʊkən 模型计量单元}}. EP64 or EP128 describes ^{{dɪˈskraɪbz 描述}} parallel ^{{ˈpærəlel 并行的}} scale ^{{skeɪl 规模}}, not the number of experts ^{{ˈekspɜːts 专家}}. Nvidia’s ^{{ɛnˈvɪdiəz 英伟达的}} NVL ^{{ˌen viː ˈel 英伟达机架系列名}} 72 is 72 GPUs ^{{ˌdʒiː piː ˈjuːz 图形处理器}}. The 2025 CloudMatrix384 paper used Ascend 910; DeepInfer’s current stated ^{{ˈsteɪtɪd 明示的}} support names Ascend 950 and NVIDIA ^{{ɛnˈvɪdiə 英伟达}} SM100. These are different units ^{{ˈjuːnɪts 单位}} and hardware ^{{ˈhɑːdweə 硬件}} generations ^{{ˌdʒenəˈreɪʃənz 代际}}. A matching ^{{ˈmætʃɪŋ 对应的}} number alone cannot prove ^{{pruːv 证明}} that the software ^{{ˈsɒftweə 软件}} was designed ^{{dɪˈzaɪnd 设计的}} specifically ^{{spəˈsɪfɪkli 具体地}} for a certain system.

[Test method ^{{ˈmeθəd 方法}} / counter ^{{ˈkaʊntə 反向的}}-evidence] The example test uses full-acceptance ^{{əkˈseptəns 接受}} accounting ^{{əˈkaʊntɪŋ 计量口径}}: decode ^{{diːˈkəʊd 解码}} speed counts 1+nextn tokens ^{{ˈtəʊkənz 模型计量单元}} per request ^{{rɪˈkwest 请求}} per step regardless ^{{rɪˈɡɑːdləs 不论}} of the true acceptance ^{{əkˈseptəns 接受}} rate. With nextn=5 and 64 requests ^{{rɪˈkwests 请求}}, that is (1+5) × 64 = 384 counted tokens ^{{ˈtəʊkənz 模型计量单元}} per step; it is not necessarily ^{{ˌnesəˈserəli 必然地}} 384 accepted ^{{əkˈseptɪd 已接受的}} output ^{{ˈaʊtpʊt 输出}} tokens ^{{ˈtəʊkənz 模型计量单元}}. Some load ^{{ləʊd 负载}}-balance ^{{ˈbæləns 平衡}} flush ^{{flʌʃ 刷新}} intervals ^{{ˈɪntəvəlz 区间}} are excluded ^{{ɪkˈskluːdɪd 排除的}}. The release ^{{rɪˈliːs 发布}} is an offline ^{{ˌɒfˈlaɪn 离线的}} inference ^{{ˈɪnfərəns 推理}} core ^{{kɔː 核心}}; a full service still needs cache ^{{kæʃ 缓存}} management ^{{ˈmænɪdʒmənt 管理}}, separated ^{{ˈsepəreɪtɪd 分开的}} prefill ^{{ˈpriːfɪl 预填充}} and decode ^{{diːˈkəʊd 解码}} stages ^{{ˈsteɪdʒɪz 阶段}} and load ^{{ləʊd 负载}} balancing ^{{ˈbælənsɪŋ 平衡}}.

13 days to use the top plan

[Official ^{{əˈfɪʃəl 官方的}} terms ^{{tɜːmz 条款}}] The current Max ^{{mæks 最高档}} list price is RMB ^{{ˌɑːremˈbiː 人民币}} 599 a month with 11,980 points. The pre-31 August allowance ^{{əˈlaʊəns 额度}} of 650 million tokens ^{{ˈtəʊkənz 模型计量单元}} became 11,980 points; model ^{{ˈmɒdl 模型}}, input ^{{ˈɪnpʊt 输入}}, output ^{{ˈaʊtpʊt 输出}} and cached ^{{kæʃt 已缓存的}} tokens ^{{ˈtəʊkənz 模型计量单元}} use different amounts ^{{əˈmaʊnts 数量}}. A large raw ^{{rɔː 原始的}} token ^{{ˈtəʊkən 模型计量单元}} count is not a cash ^{{kæʃ 现金}} bill ^{{bɪl 账单}}, a new price rise, or the same number of paid output ^{{ˈaʊtpʊt 输出}} tokens ^{{ˈtəʊkənz 模型计量单元}}.