[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"news-slug-qwen3-8-2-4t-trillion-cowork-agent-launch":3},{"id":4,"title":5,"summary":6,"content":7,"original_url":8,"source_id":9,"tags":10,"translations":24,"news_slug":31,"published_at":32,"created_at":33,"modified_at":34,"is_published":35,"publish_type":36,"image_url":14,"view_count":37},"5c6d9de5-aef2-4490-96f1-e82166cc44ec","阿里千问 Qwen3.8 正式发布：2.4T 参数的旗舰基座，首次把 Cowork 塞进 Agent 入口","阿里巴巴于 8 月 3 日正式发布新一代基座大模型 Qwen3.8，总参数量达 2.4 万亿，主打编程与专业办公（Cowork）场景。当日 LMArena 三方榜中，阿里 Qwen 系列位列 Anthropic Claude 系列之后排第二；API 已同步上线千问 AI 平台，Qwen3.8-Max 与 Qwen3.8-27B 预计下周开源。","## 阿里千问 Qwen3.8 正式发布：2.4T 参数的基座旗舰，Agent 第一次被推到产品主位\n\n8 月 3 日，阿里巴巴正式发布新一代基座大模型 **Qwen3.8**，总参数量达到 **2.4 万亿**。发布当天，Qwen3.8 的 API 已上线千问 AI 平台，并直接接入同一天推出的企业级 Agent 产品“**千问办公（QwenWork）**”——网页版与独立 PC 客户端率先开放，钉钉 PC 端与手机端内置入口将在近期上线 [1]。\n\n这一次的发布节奏里隐藏着一个比参数更重要的信号：**阿里把基座模型和 Agent 产品首次做成了同日发布**。同一个工作日，模型上线、Cowork 场景落地、Agent 产品入口全开——这并不是把模型塞进一个聊天框那么简单。\n\n### 旗舰能力集中在两类工作：Coding 与 Cowork\n\n按照官方口径，Qwen3.8 在 **编程（Coding）** 和 **专业办公（Cowork）** 两个方面做了重点强化 [1]。从 7 月 19 日开放的 **Qwen3.8-Max-Preview** 内部测评看，其面向“全栈开发、数据分析、Office 办公工作流”等多 Agent、长程复杂任务的能力，官方表述为“匹敌世界前沿模型的实力水平” [2]。\n\n这并不是单纯的 C 端写作能力升级——它把目标工作流压到了**长程、多步骤、跨工具** 这一档，对应的真实场景是开发者在 IDE 里的长任务、企业员工在 Office\u002F数据分析里跑出的多步骤协作。\n\n### LMArena 三方榜，阿里排第二，排第一的是 Anthropic Claude\n\n发布当日，**LMArena 三方榜单**放榜，阿里 Qwen 系列的排名位列 Anthropic Claude 系列之后，**居第二** [1]。在 2.4T 总参数规模下，排在闭源的 Claude 系列之后，这一位置对开源\u002F半开源路线来说已经是头部成绩。\n\n需要注意的是，这里的对比锚点是**三方榜单的总排名**，而不是某个具体的 benchmark 跑分。文章对应到的是 8 月 3 日当日的榜单口径，而不是某一周内的滚动数据。\n\n### 开源节奏下周落地：Qwen3.8-Max + Qwen3.8-27B\n\n发布通稿明确：Qwen3.8-Max **预计下周开源**，**同时还将开源 Qwen3.8-27B** [1]。这意味着对独立开发者和小团队来说，**真正能本地化部署、或者在云上微调的版本要再等几天**。\n\n加上千问 AI 平台已经在跑的 API，可以理解为阿里在同一天铺了三件事：把旗舰模型给到大客户和 Agent 产品、把 API 给到开发者、把“即将开源”的时间点给到社区。三种触点在 24 小时内一起动，节奏比之前几代都密。\n\n### 为什么“2.4T 参数”这次不是单独卖点\n\n参数堆叠一直是国内基座模型的标配叙事，这次的不同在于它**没有像过去那样单独拉出来宣传**。在这篇 36氪的发布通稿里 [1]，“2.4 万亿参数”只是定语，**真正作为分段落核心卖点的是 Coding 与 Cowork 两个场景**。\n\n这种“从参数规模转向场景能力”的叙事切换，反映出基座模型竞争已经从“谁的总参数大”过渡到“谁能把多 Agent 长程工作流跑通”。如果只看 LMArena 总榜，Qwen3.8 仍在 Anthropic 之后；但配合同日上线的 QwenWork Agent 产品，阿里其实在押一条**“模型 + Agent 入口”一体化**的路。\n\n### 所以呢：对开发者意味着什么？\n\n短期来看，做选择的不是模型大小，而是**接入路径**：\n\n- **想用 Qwen 3.8 的 API**：今天就能在千问 AI 平台上跑；\n- **想做企业 Agent 集成**：盯“千问办公”开放节奏——网页\u002FPC 客户端现已上线，钉钉入口近期落地；\n- **想本地化或者私有云**：等下周开源的 Qwen3.8-Max 和 Qwen3.8-27B，这是真正能跑微调和部署的版本。\n\n如果 Qwen3.8-Max 在开源后的复现跑分基本对齐官方口径，加上同源的 QwenWork Agent 入口，这一代 Qwen 走出来的不是“又一个大模型”，而是**“把大模型 + Agent 产品 + 开源分叉”三条线在一个工作日打通**的组合拳。\n\n---\n\n**引用**\n\n[1] 36氪：阿里巴巴正式发布新一代基座大模型 Qwen3.8，总参数量 2.4 万亿 — https:\u002F\u002F36kr.com\u002Fnewsflashes\u002F3923193948614025\n\n[2] IT之家：阿里千问 Qwen3.8 预览版模型上线：2.4T 参数量，宣称“可能是除了 Fable 5 外最强大的模型” — https:\u002F\u002Fwww.ithome.com\u002F0\u002F978\u002F730.htm","https:\u002F\u002F36kr.com\u002Fnewsflashes\u002F3923193948614025","5e4fd3d1-9cb4-44a6-bae5-9ffb449c05c1",[11,15,18,21],{"id":12,"name":13,"slug":13,"description":14,"color":14},"a8002d98-9df1-4ab9-94d4-a7625af634c4","china-ai",null,{"id":16,"name":17,"slug":17,"description":14,"color":14},"01598627-1ea6-4b27-a5d8-874971571a71","llm",{"id":19,"name":20,"slug":20,"description":14,"color":14},"7e89b5cc-57db-4f37-bc6d-28919a73931c","model-release",{"id":22,"name":23,"slug":23,"description":14,"color":14},"c187600e-804c-4697-b828-1e4330e0eb10","qwen",[25],{"id":26,"lang":27,"title":28,"summary":29,"content":30},"099d32d2-4056-4b5d-a087-aa258c3fcb86","en","Alibaba's Qwen3.8 Officially Released: 2.4T-Parameter Flagship Base Model Brings Cowork into the Agent Entry Point","On August 3, Alibaba officially released its new flagship base model Qwen3.8 with 2.4 trillion total parameters, focusing on Coding and professional office (Cowork) workflows. On the same-day LMArena third-party leaderboard, Alibaba's Qwen series ranked second behind Anthropic's Claude series; the API is now live on Qianwen AI, with Qwen3.8-Max and Qwen3.8-27B expected to open-source next week.","## Alibaba's Qwen3.8 Officially Released: A 2.4T Flagship Base Model That Puts Cowork Into the Agent Entry Point\n\nOn August 3, 2026, Alibaba officially released its new flagship base model **Qwen3.8**, with a total parameter count of **2.4 trillion**. On release day, the Qwen3.8 API went live on the Qianwen AI platform and was directly wired into the same-day-launched enterprise Agent product **QwenWork** (\"Qianwen Banhui\") — the web version and a standalone PC client opened first, with the DingTalk PC and mobile embedded entry points slated to come online shortly [1].\n\nBeneath the parameter headline there is a more important signal: **Alibaba for the first time launched the base model and an Agent product on the same day**. Model go-live, Cowork scenario shipping, and Agent entry points all landed in a single working day — which is not simply a model dropped into a chat box.\n\n### Two flagship scenarios: Coding and Cowork\n\nAccording to the official announcement, Qwen3.8 was deliberately tuned for **Coding** and **professional office workflows (Cowork)** as its two focal tracks [1]. The internal evaluation of the **Qwen3.8-Max-Preview** opened on July 19 describes its capability on multi-Agent, long-horizon, complex tasks such as full-stack development, data analysis, and Office workflows as \"matching world-leading models\" [2].\n\nThis is not a C-end writing upgrade — the target workflow has been pushed into the **long-horizon, multi-step, cross-tool** tier. The real scenarios in mind are developers' long tasks inside the IDE and employees' multi-step collaboration across Office and data analysis tools.\n\n### LMArena third-party leaderboard: Alibaba ranks second, behind Anthropic Claude\n\nOn release day, the **LMArena third-party leaderboard** was refreshed; Alibaba's Qwen family ranked behind Anthropic's Claude family, **placing second** [1]. At a total scale of 2.4T parameters, sitting just behind the closed-source Claude family is already a top-tier position for the open \u002F semi-open track.\n\nCaveat: the comparison anchor here is the **overall ranking on the third-party leaderboard**, not a specific benchmark score. The figure refers to the leaderboard snapshot on August 3, not a rolling weekly aggregate.\n\n### Open-source timeline: Qwen3.8-Max + Qwen3.8-27B landing next week\n\nThe release announcement is explicit: **Qwen3.8-Max is expected to open-source next week**, **and Qwen3.8-27B will also open-source at the same time** [1]. For independent developers and small teams, **the versions that can actually be self-hosted or fine-tuned in a private cloud will have to wait a few more days**.\n\nCombined with the API already running on Qianwen AI, Alibaba laid down three things on the same day: a flagship model for big customers and the Agent product, an API for developers, and an \"about-to-open-source\" timeline for the community. All three touch points moved within a 24-hour window — a denser cadence than the previous generations.\n\n### Why \"2.4T parameters\" is not the sole pitch this round\n\nParameter scale has long been the default headline for Chinese base models. The difference this time is that **the announcement did not single it out for separate promotion**. In the 36kr release write-up [1], \"2.4 trillion parameters\" is a modifier — **the actual core pitch by paragraph is the two scenarios: Coding and Cowork**.\n\nThis narrative shift — from parameter scale to scenario capability — shows that base-model competition has moved past \"who has the bigger total parameter count\" to \"who can actually run multi-Agent, long-horizon workflows end-to-end\". Judged by the LMArena overall ranking alone, Qwen3.8 is still behind Anthropic; paired with the same-day QwenWork Agent product, however, Alibaba is in fact betting on a **\"model + Agent entry point\" integrated** play.\n\n### So what: what does it mean for developers?\n\nIn the short term, the variable is not model size — it is the **access path**:\n\n- **Want to call Qwen 3.8 via API**: live today on the Qianwen AI platform;\n- **Want enterprise Agent integration**: watch the QwenWork rollout — web\u002FPC clients are live now, DingTalk entries are landing shortly;\n- **Want self-hosting or private cloud**: wait for next week's open-source Qwen3.8-Max and Qwen3.8-27B, which are the versions that actually enable fine-tuning and deployment.\n\nIf Qwen3.8-Max's post-open-source reproducible benchmarks broadly match the official claims, combined with the same-source QwenWork Agent entry point, this generation of Qwen is not \"yet another LLM\" — it is a **three-line combination \"big model + Agent product + open-source fork\" all wired up in a single working day**.\n\n---\n\n**References**\n\n[1] 36kr: Alibaba Officially Releases the New-Generation Base Model Qwen3.8, Total Parameter Count 2.4 Trillion — https:\u002F\u002F36kr.com\u002Fnewsflashes\u002F3923193948614025\n\n[2] IT之家 (ITHome): Alibaba Qwen Qwen3.8 Preview Model Online: 2.4T Parameters, Claims to Be \"Perhaps the Most Powerful Model Apart from Fable 5\" — https:\u002F\u002Fwww.ithome.com\u002F0\u002F978\u002F730.htm","qwen3-8-2-4t-trillion-cowork-agent-launch","2026-08-03T08:30:00Z","2026-08-03T08:04:35.176598Z","2026-08-03T08:04:35.176607Z",true,"agent",252]