美团用CatPaw免费送GLM-5.2给程序员,阿里用QoderWork免费送GLM-5.2给职场人。两家大厂几乎同时做了同一件事,但面向完全不同的人群。
这背后藏着一个正在成型的趋势:大厂不再只做模型,而是把顶级模型塞进具体工作场景里做"工具+模型"的捆绑分发。
先说产品。阿里云的通义灵码在今年5月20日改名成了Qoder CN,整个产品线做了一次大重构:编码的归编码,办公的归办公。其中面向日常办公场景的子产品叫QoderWork,定位是"桌面端AI工作搭子"——不是帮你查资料回答问题,是直接帮你把活儿干了。
但真正值得聊的不是产品本身,而是它折射出的行业信号。
一、GLM-5.2:大厂入口之争的"通用弹药"
QoderWork的模型选择器里,列了五款模型:Qwen3.7-Max(0.25X)、Qwen3.7-Plus(0.1X)、DeepSeek-V4-Pro(0.5X)、DeepSeek-V4-Flash(0.1X),以及GLM-5.2(0.6X)。
GLM-5.2消耗积分最高,但也是里面能力最强的编程模型。继美团CatPaw之后,阿里QoderWork也把它塞进来了。
为什么大厂都选它?因为GLM-5.2有三个特点:足够强(Code Arena全球第一)、足够便宜(开源免费)、无限制(MIT协议,谁都可以拿来商用)。这三个条件叠加,让它成了这场入口之争里的通用弹药。
二、QoderWork能干什么?四个核心能力
能力1 | 连接器:打通钉钉、飞书、Notion、Figma
可以对接钉钉、飞书、Slack、Microsoft 365、Notion、Figma、Canva等外部软件。你让AI在钉钉群自动回复、在飞书文档查数据、在Notion更新知识库、在Figma拉设计稿——不需要手动切窗口,直接调用对方API。
能力2 | 定时任务:给AI设闹钟,到点自己干
每天早9点半自动读Excel生成日报,每天下午6点半自动清理下载文件夹归档,每周一10点自动追踪竞品动态。人不在电脑前,它按点执行。之前这种功能只在Zapier、n8n等企业级自动化平台里见过。
能力3 | IM频道:钉钉飞书微信直接调用
接入钉钉、飞书、Lark、微信、企业微信的消息通道,在IM里@机器人就能跑任务、查数据、整理文档,结果直接回到聊天窗口。可以搭AI客服、团队自动应答机器人、群消息摘要助手。
能力4 | 专家套件广场:按行业角色预装技能
行政、HR、财务、律师、市场、运营等十几个类目,每个套件预装了行业逻辑和工作流程。比如合同管理套件能审查风险、起草初稿、红线对比;咨询交付套件覆盖从桌面调研到CEO汇报七个场景。
三、免费背后:积分制和真实成本
新用户注册自动获得300积分,14天内可领1700,加起来2000。学生教师认证再多送4000。首月直接是Pro专业版(正常价59元/月),不用付钱。
但实测下来,一个简单问题用GLM-5.2就吃掉4个积分,占套餐的2%。2000积分按0.6X消耗跑GLM-5.2,大概几百轮对话就见底。积分有效期只有30天,没用完就作废。
使用策略建议:日常简单任务挂Qwen3.7-Plus(0.1X),能跑6倍对话量;只有遇到真正需要强推理的复杂任务,才切到GLM-5.2。杀鸡不用牛刀。
四、智谱的"安卓策略"
把视角拉高,你会发现一个更大的判断正在浮现。
美团用CatPaw免费送GLM-5.2给程序员,阿里用QoderWork免费送GLM-5.2给职场人,之前美团还有个Tabbit浏览器免费送GPT-5.5。大厂们正在做"工具+模型"的捆绑分发——浏览器、IDE、办公桌面,每个入口都在争夺用户的使用习惯。
而GLM-5.2这个开源模型,正在成为这场入口之争的通用弹药。智谱的策略到这里就很清楚了:它不跟你争终端用户,它做弹药供应商。模型免费开源,MIT协议无限制,让所有产品来分发。
我的判断
这个逻辑跟安卓当年很像。谷歌不靠安卓系统本身赚钱,它靠安卓成为所有手机的底层操作系统之后,用搜索、广告和应用商店变现。智谱也不需要靠GLM-5.2的API调用费赚大钱,它需要的是让这个模型变成中国AI工具生态里的默认选项。一旦所有主流工具里都跑着GLM-5.2,智谱在这个生态里的话语权就稳了。免费开源一个模型,看起来是放弃了收入。但如果这个模型成了所有工具的底层标配,那它就成了基础设施。基础设施的价值,从来不是按流量收费算出来的。
Meituan gave away GLM-5.2 for free to programmers via CatPaw, and Alibaba gave away GLM-5.2 for free to office workers via QoderWork. Two major companies did almost the same thing at the same time, but targeting completely different audiences.
This reveals a forming trend: large companies are no longer just developing models, but are embedding top models into specific work scenarios to bundle and distribute them as “tools + models”.
First, the product. Alibaba Cloud's Tongyi Lingma was renamed Qoder CN on May 20 this year, and the entire product line underwent a major reconfiguration: coding goes to coding, and office work goes to office work. The sub-product aimed at daily office scenarios is called QoderWork, positioned as a "desktop AI work buddy"—not to look up information or answer questions for you, but to actually get the work done.
But the real story here isn’t just about the product — it’s about the signal it sends to the industry.
**I. GLM-5.2: The “Universal Ammo” for Big Company Entry Wars**
QoderWork’s model selector lists five models: Qwen3.7-Max (0.25X), Qwen3.7-Plus (0.1X), DeepSeek-V4-Pro (0.5X), DeepSeek-V4-Flash (0.1X), and GLM-5.2 (0.6X).
GLM-5.2 consumes the most points but is also the most powerful programming model. Following Meituan’s CatPaw, Alibaba’s QoderWork has also included it.
Why do big companies choose it? Because GLM-5.2 has three key features: it’s **powerful enough** (ranking first in the Code Arena global competition), **cheap enough** (open-source and free), and **unrestricted** (MIT license, allowing commercial use by anyone). These three conditions make it the universal ammo for this battle for market entry.
**II. What Can QoderWork Do? Four Core Capabilities**
**Capability 1 | Connector:** Integrates with DingTalk, Feishu, Notion, Figma
It can connect to external software like DingTalk, Feishu, Slack, Microsoft 365, Notion, Figma, and Canva. You can instruct the AI to automatically reply in a DingTalk group, check data in a Feishu document, update a knowledge base in Notion, or pull a design draft from Figma — all without switching windows, just by calling the API.
**Capability 2 | Scheduled Tasks:** Set an alarm for AI; it runs on its own
Automatically read an Excel file to generate a daily report at 9:30 AM, clean up and archive the download folder at 6:30 PM, or track competitor dynamics every Monday at 10 AM. Even when you’re not at your computer, it executes tasks on schedule. Previously, this level of automation was only seen in enterprise automation platforms like Zapier or n8n.
**Capability 3 | IM Channels:** Direct invocation from DingTalk, Feishu, WeChat
Connects to messaging channels like DingTalk, Feishu, Lark, WeChat, and WeChat for Business. You can run tasks, check data, or organize documents by tagging the robot in an instant message, with the results appearing directly in the chat window. This enables AI customer service, team auto-response bots, and group message summary assistants.
**Capability 4 | Expert Kit Marketplace:** Pre-installed skills by industry role
Includes dozens of categories such as administration, HR, finance, lawyers, marketing, and operations. Each kit pre-loads industry-specific logic and workflows. For example, the contract management kit can review risks, draft initial versions, and perform redline comparisons; the consulting delivery kit covers seven scenarios from desktop research to CEO reports.
**III. Behind the Free Offer: The Point System and Real Cost**
New users automatically receive 300 points upon registration, with an additional 1700 available for free over 14 days, totaling 2000. Students and teachers can get an extra 4000 points. The first month is free and includes the Pro plan (normally 59 RMB/month).
However, in practice, a simple question using GLM-5.2 costs 4 points, which is 2% of the package. With 2000 points at a 0.6X consumption rate for GLM-5.2, you can only expect a few hundred conversations before it runs out. Also, the points expire after 30 days.
**Usage Strategy:** For everyday tasks, stick with Qwen3.7-Plus (0.1X), which offers six times the conversation volume for the same points. Only switch to GLM-5.2 for complex tasks that require strong reasoning. No need to use a cannon to kill a mosquito.
四、智谱的"安卓策略"
When you zoom out, you see a larger trend emerging.
Meituan’s CatPaw and Alibaba’s QoderWork are both distributing GLM-5.2 for free, and previously, Meituan offered GPT-5.5 for free in the Tabbit browser. Big companies are pushing for **“tool + model” bundled distribution** — browsers, IDEs, office desktops, every entry point is vying for user habits.
GLM-5.2, as an open-source model, is becoming the universal ammo in this fight. Zhipu’s strategy is clear: **it’s not competing for end users; it’s positioning itself as the ammo supplier.** The model is free and open-source under MIT, allowing any product to distribute it.
我的判断
This logic is very similar to Android. Google didn't make money from the Android system itself; it made money by using Android as the underlying operating system for all smartphones, monetizing through search, advertising, and the app store. Zhipu doesn't need to make big money from GLM-5.2 API call fees either; what it needs is for this model to become the default choice in China's AI tool ecosystem. Once all mainstream tools run on GLM-5.2, Zhipu's influence in this ecosystem is secured. Making a model free and open source may seem to give up income. But if this model becomes the standard for all tools, it becomes infrastructure. The value of infrastructure is never calculated by charging for traffic.
#QoderWork#GLM-5.2#阿里云#智谱#AI办公