Qianwen Office

• 1. 概要 (318 chars)

  千问办公（QwenWork）是阿里巴巴于 2026 年 7 月 27 日上线的一站式 AI Agent 办公平台，由钉钉 CEO 陈宇森牵头，整合 QoderWork、悟空、MuleRun 三款内部 Agent 产品而成，官网 qwenwork.cn 同步开放 Windows 与 macOS 桌面客户端公测。产品以 Qwen3.8-Max-Preview 模型为底座，主打「对话即交付」——用户用自然语言描述需求，系统直接产出 Word、PPT、Excel、数据分析报告乃至可发布的网页。媒体实测认可其 PPT 生成与长文档处理能力突出，但生成速度、跨应用兼容性与数据幻觉问题仍是短板。它被普遍视为阿里在 AI 办公入口之战中对标腾讯 WorkBuddy、字节 TRAE、金山 WPS 的核心武器。

• 2. 产品背景 (494 chars)

  千问办公的诞生带有明显的「整合止战」色彩。此前阿里内部同时孵化了三款 Agent 产品：源自阿里云体系的桌面级智能体 QoderWork、由钉钉前 CEO 陈航主导的企业级 AI 工作平台「悟空」、以及 2025 年内部孵化、主攻海外的跨平台 Agent 执行引擎 MuleRun。2026 年 6 月 11 日陈宇森接任钉钉 CEO 后，在 46 天内结束内部赛马，一周内推动悟空与 MuleRun 合并，7 月初再纳入 QoderWork，形成「前台×中台×后台」的完整架构，最终以「千问办公」之名统一亮相。   界面新闻评价称，阿里「不等所有产品准备好就上线，足见腾讯给阿里的压力之大」——据易观数据，2026 年 6 月腾讯系智能体访问量约为阿里系的 3.5 倍，其中腾讯 WorkBuddy 单月访问量突破 2000 万，而 QoderWork 仅约 788 万。千问办公被定位为独立产品而非钉钉子功能，但采取「独立+嵌入」双形态：既可在桌面端独立运行，也能在钉钉侧边栏直接调用，背靠钉钉超 2600 万家企业组织的生态。需要注意的是，千问办公与面向 C 端的千问 App（通义千问 App）是两个不同的产品，后者的 AI PPT、会议纪要等功能早在 2025 年 12 月就已上线。

• 3. 产品功能与使用体验 (534 chars)

  千问办公的核心逻辑是「任务对话交付」：用户以自然语言描述需求，系统通过多轮对话拆解并执行复杂任务，直接交付成品。官方文档列出的能力覆盖 Office 全链路生成与编辑（Word、Excel、PPT、PDF、HTML 原生预览与实时编辑）、AI PPT 幻灯片工作台、数据分析与可视化、网页全栈生成与发布（内置 Pages 托管，可生成带数据库和交互逻辑的免部署网页）、多模态理解与创作，以及打通钉钉群聊、考勤审批、会议日程、知识库、邮件等约 25 项企业协作能力。桌面端还提供电脑操控、浏览器自动化、应用快照、Hooks、云端定时任务与跨会话记忆等进阶功能。   实测表现方面，智东西用 7 分钟生成了一份 12 页的 Alphabet 财报 PPT，包含 11 个可编辑图表，核心数据与官方财报吻合；2 分钟完成播客总结并推送到飞书文档。36 氪爱点评的用户反馈称长文档处理是「杀手锏」，上传研报、合同、政策文件数秒即出结构化摘要，一小时的工作缩到几分钟。但短板同样明显：飞书文档转换出现文字乱码导致任务失败，电商网页生成耗时 1 小时 8 分钟，60 页 PDF 转换只完成约六成，股票实时数据首次生成出错且不自知。网易的评测以「六边形战士还是缝合怪」为题，认为结构化任务接近可用，但实时数据校验与跨模态融合「表现不及格」。

• 4. 定价与商业模式 (339 chars)

  千问办公采用免费加订阅的三层结构。个人免费版 0 元并限时赠送积分；个人标准版 98 元/月（当前 8 折 78 元）；个人高级版 198 元/月（8 折 158 元）；企业标准版 198 元/月/席；企业旗舰版走定制化咨询，支持更严格的网络与部署要求。会员可调用基础与高级模型，享高速通道不限量对话，高档位对应更多积分、任务数与云盘容量。企业版在阿里云官网以「席位+积分包」形式购买。   值得注意的是积分消耗速度：智东西单次实测消耗约 1000 积分，相当于个人标准版半个月的额度；有重度用户反馈 3 天就触发额度提醒。商业化前景方面，界面新闻指出企业客户更关心任务完成率、数据安全、权限控制与结果留痕，而微软 Copilot 的前车之鉴——4.5 亿商业客户中完整版付费比例不足 4.5%——提示「模型能力不等于落地能力」。

• 5. 用户评价 (378 chars)

  正面评价集中在几个方面。长文档处理被多位用户称为最大亮点，研报拆解、合同比对、政策文件提炼的效率提升立竿见影。中文正式文案质量高，商务邮件与公文「改一改就能直接用」，在公众号、小红书、短视频三种文体的风格区分上被评为同类产品中最自然。电商工作流有实际落地价值，有用户用 1688 技能完成手机壳爆款进货攻略，给出热度评级与利润区间。三合一整合后权限开放程度高于早期沙箱化的悟空，Computer use 与 Browser use 对普通用户有真实价值。   负面反馈同样具体。跨应用兼容性问题突出，飞书文档转换乱码、表格复制到 Word 或 WPS 后格式错乱；生成速度偏慢，网页与 3D 任务常常最慢；多轮对话超过 5 轮后出现「失忆」，推翻此前结论；数据幻觉问题让结果必须人工复核，三款竞品对同一票房数据给出 85%、92%、80% 三个打架的结果且均无统计口径；内测期还偶发任务莫名中断。

• 6. 行业评价 (310 chars)

  媒体共识是 AI 办公竞赛已从「谁更能聊」转向「谁更能执行」。AI 蓝媒汇认为千问办公补上了阿里桌面 Agent 的短板，打通了从任务到交付的全链路。在 AIX 财经的六轮三产品对比实测中，千问办公的复杂生成完成度与成品质量拿到最高分，但「愿意花更多时间补细节，效率被拖慢，首轮交付不稳定」。竞争格局上，腾讯 WorkBuddy 7 月日活突破 1300 万，办公理解与主动补充观点能力强；字节 TRAE 工程结构清晰、速度快，但常把交付做成网页，对传统办公格式「水土不服」；金山 WPS 走文档原生加私域数据路线，凭 6.78 亿月活设备的数据壁垒被称为「独一份的答卷」。中信证券的观点被多家媒体援引：竞争关键点正从「模型选型」转向「用好企业私域数据」。

• 7. 争议与风险 (252 chars)

  最大的争议在于数据隐私。千问办公需要授权读取群聊、考勤、审批、邮件等企业数据，36 氪复盘钉钉此前 ONE 项目的失败——上线后次日留存从 45% 跌到 18%，超 60% 用户三次使用即卸载，差评关键词直指「监视、内卷、压迫」——警示监控式入口的路线风险。其次是整合风险，三款 Agent「强行合体」被质疑为缝合怪，原产品的权益、数据与团队融合深度尚待验证。幻觉问题对企业级交付尤为致命，模型「擅长生成答案、不擅长判断」，结果无法被完全信任。此外，阿里系流量显著落后于腾讯系，窗口期短，能否在红海竞争中站住脚仍是未知数。

• 8. 适用人群与使用建议 (310 chars)

  千问办公适合三类人群：一是需要高频处理长文档的专业岗位，如律师法务的合同比对、投研分析师的年报拆解、教师的课件制作；二是深度使用钉钉的企业组织，可以直接在钉钉侧边栏调用，打通财务、人事、市场、运营全流程；三是电商与自媒体从业者，数据复盘与营销方案生成有现成技能可用。学生写论文、求职者改简历等轻量场景用免费版即可。   不适合的场景包括：对实时数据准确性要求高的任务（股票行情等仍有幻觉），重度依赖飞书或 WPS 生态的团队（跨应用兼容性差），以及对生成速度敏感的即时协作场景。如果核心需求只是文档内的 AI 辅助，WPS 的文档原生方案学习成本更低；如果追求生成速度，字节系产品更快。重度用户需留意积分消耗速度，标准版额度可能不够用。

• 9. 总结 (136 chars)

  千问办公是阿里放弃内部赛马、以整合姿态押注 Agent 办公的关键一役——产品完成度已能交付真实价值，但速度、稳定性与隐私信任仍是其能否在腾讯、字节、金山的围攻中站稳的三道坎。随着 Qwen 模型持续迭代与钉钉生态深度打通，它有机会成为国内企业级 AI 办公的头部入口，值得持续关注。

No in-depth report yet

• 

  uedgerLion171

  —

  Has anyone else run into tables getting mangled when copied into Word or WPS? I have to manually fix the formatting every single time. So annoying.

---

• 

  WPetersonQ4

  —

  Honest thoughts after three days of use: long-document processing is genuinely impressive. Drop in a few-hundred-page research report and you get a structured summary in seconds, and it spots contradictory clauses accurately too. Work that used to take an hour now takes minutes. But it really burns through credits — as a heavy user I hit the quota warning on day three, and the standard plan allowance feels nowhere near enough.

---

• 

  松涛

  —

  It burns credits way too fast, I can't keep up!!

---

• 

  smallbutterfly659

  —

  Lawyers are already using it for contract comparison, and industry summits are even handing out Wukong invite codes to promote the legal agent. The speed of industry adoption is a little scary.

---

• 

  VRiveraIII

  —

  The IM channel can hook into WeChat and DingTalk messages, letting the agent join your daily chats like a real person. Pretty novel gameplay, though it feels a bit uncanny when you think about it.

---

• 

  BrendaCookII

  —

  Way more polished than I expected.

---

• 

  KBrooks_20244

  —

  DingTalk has QwenWork built right in — you can summon it from the sidebar and handle group messages and schedules without switching windows. Pretty sweet for people who live in DingTalk all day.

---

• 

  Gary.Diaz00740

  —

  The desktop client has the fullest feature set — browser automation and local file operations are client-only. The web version is just a lightweight entry point; for real work, just download the desktop app.

---

• 

  JerryRajirez

  —

  The free tier gives 2000 initial credits plus 100 daily, which is enough for making PPTs, writing docs and looking things up. Heavy users will need the 78-yuan-a-month standard plan with 2000 monthly credits plus 200 daily and 10GB of cloud storage. Heads up: daily credits expire the same day on every tier, so get in the habit of using them up. Also, flagship models on complex agent tasks drain credits fast — warm up with simple tasks before going heavy.

---

• 

  SusanHoward52026

  —

  I asked it to mimic Apple's official site and build an intro page for the iPhone 18 Pro. The page structure and animations looked surprisingly legit. This kind of work used to require scheduling a front-end dev; now it's one sentence.

---

• 

  JoeAdamsJr972

  —

  Honestly hesitant about the privacy side. It asks for access to group chats, attendance records, approvals and even email. Efficiency is up, but it feels like walking around exposed.

---

• 

  侠客_6

  —

  Open the interface and it instantly smells like Codex. Anyone who has used similar products can pick it up seamlessly.

---

• 

  xJacintoBallesteros_88

  —

  I had it write posts in three formats — WeChat official account, Xiaohongshu, and short video scripts. Its style differentiation is the most natural among all the products I tried. Copywriters should give it a shot.

---

• 

  JHenderson_77

  —

  DingTalk users are absolutely thrilled this time.

---

• 

  Jeffrey.Hart_7

  —

  After comparative testing of three AI office apps, Qwen's completion rate and output quality on complex generation are indeed the highest, but it likes to spend extra time polishing details, which drags down efficiency, and first-round deliveries are often shaky — you have to push it a few rounds to get satisfying results. Also, the three products gave conflicting figures of 85%, 92% and 80% for the same box-office stat, none with methodology stated, so whichever you use, always manually verify key data.

---

• 

  ShirleyStephens_99

  —

  The enterprise plan is 198 yuan per seat. Not exactly cheap, not exactly expensive. What matters is task completion rate, data security and permission control — bosses will probably wait and see for a while before jumping in.

---

• 

  greenswan885

  —

  Tested AI spreadsheet analysis and PPT generation — no real problems. The working-directory design is quite friendly for beginners just getting started with AI agents.

---

• 

  猫咪_3

  —

  The quality of formal Chinese copy is genuinely high. Business emails and official documents come out logically rigorous with a natural tone — basically ready to send after light edits.

---

• 

  Amy.CoxZ

  —

  Tasks kept dying mysteriously during the beta — took two reruns to succeed. Hoping the stable release is more solid.

---

• 

  Roy_Rogers_7

  —

  I asked it to organize my computer desktop. Instead of diving right in, it asked me three questions first: what logic to sort files by, how to handle shortcuts, and whether to delete temp files. That awareness of when to confirm makes it far more reliable than agents that just charge ahead blindly — real intelligence isn't just executing commands, it's knowing when to check with the user. Ten minutes later, done, and my desktop was genuinely clean. In this respect it's close to a dependable human assistant.

---

• 

  William_Anderson_Pro04

  —

  Its system design still lags Codex, but stitching Qoder's CLI capabilities together with Wukong's workflows, plus more open permissions than the early sandboxed Wukong, actually makes it more practical for ordinary users.

---

• 

  Judith43_fi

  —

  So is it a Frankenstein patchwork or a hexagonal warrior after all?

---

• 

  罗凤昊

  —

  As a longtime Wukong user I'm torn about migrating. From my comparison, Wukong is a DingTalk-native agent tied to that ecosystem, while QwenWork is a dual-form product — standalone desktop plus DingTalk embed — with a fuller feature set, including computer control and browser automation on desktop. But the new product still needs time to mature, and officials haven't clarified how existing data and benefits migrate. Running real business on it, I don't dare switch rashly. Watching from the sidelines until the stable release.

---

• 

  happyelephant354

  —

  QoderWork plus Wukong plus MuleRun, three-in-one, running on Qwen3.8-Max-preview. Alibaba is clearly done with internal horse-racing this time.

---

• 

  organicbutterfly946

  —

  One round of testing burned 1000 credits, and the standard plan only gives 2000 base credits a month. That's half a month's allowance gone. This pricing stings a bit.

---

• 

  掠影_14

  —

  The web page generation task ran for a full hour and eight minutes. The interaction and UI are genuinely nice — cart, checkout and coupons all included — but that speed is just unbearable.

---

• 

  CStewart_77

  —

  Feishu document conversion output pure gibberish and the task failed. Cross-app compatibility still needs work.

---

• 

  JButler168_391

  —

  It generated a 12-page earnings PPT in 7 minutes with 11 editable charts. I cross-checked the core numbers against Alphabet's official report and they basically all matched — revenue up 24%, Google Cloud up 82%, net profit up 298%. Clean structure, consistent color scheme, totally presentation-ready. It even does deep thinking first, splitting the task into six subtasks and executing them one by one. The whole workflow feels methodical. This level of completeness is seriously something.

---

• 

  Henry.CastilloIII

  —

  It flopped on same-day tech stock quotes. The first batch of data was just wrong, and after several rounds of back-and-forth it still didn't notice the problem until I threw a screenshot from my trading app in its face. The old hallucination problem remains unsolved — always double-check anything involving real-time data.

---

• 

  TheAleksiLinna

  —

  Fed it a podcast transcript and it distilled the key points and quotable lines in 2 minutes, then, after authorization, wrote them straight into a Feishu doc and sent it. That's a smooth pipeline.

No reviews yet