国产大模型三国杀:百度搭子、美团 LongCat、腾讯 Hy3,打工人该把活交给谁?Chinese AI Trio Face-Off: Baidu, Meituan, and Tencent — Which Model Should You Trust with Your Workload?


AI free token/AI免费积分

本文来源:https://mp.weixin.qq.com/s/e4YPJ61y7qGq6hxZ5dBVgw
若有版权问题,请来信告知:13951431913#139.com 或在右方提交评论。

这周国产 AI 圈有点不寻常。

百度、美团、腾讯,三家在同一周拿出能"干活"的大模型。不是又发个聊天机器人刷存在感,是实打实跟你喊:"我帮你把活干完。"

但打工人早就过了"看参数就激动"的阶段。你我只关心一句话:

谁真的能替我搬砖,还别太贵?

今天把这三个"国产 AI 员工"摆在一起,不吹不黑,直接给结论。

图片

一、腾讯 Hy3:微信里就能用的"免费 AI 员工"

7 月 6 号发布的混元 Hy3,是这三个里最容易上手的一个。

说人话:它是一个专门为"干活"设计的模型,不是聊天,不是吟诗,而是帮你处理真实办公任务——做 PPT、写 Excel 分析、生成网页、回知识库问题。总参数 2950 亿,但每次干活只激活 210 亿(约 7%)。

它的杀手锏不是参数,是入口。

  • 元宝(微信小程序):接入 Hy3 后多了 Agent 功能。你说"帮我做一份行业分析报告,搜最新数据,出 PPT",它会自己上网搜、交叉验证、做分析、画图表,最后直接丢给你一个 PPT 文件。而且——全部免费。
  • WorkBuddy(企业微信里):接入 Hy3 后任务成功率从 72% 跳到 90%,平均耗时缩短三分之一。会议纪要、数据清洗、图表绘制,就在你的工作流里。
  • 微信生态:公众号 AI 分身意图识别准确率 98.94%,已经悄悄覆盖了 10 亿+ 用户。

适合谁:只想在微信里说句话就把活干了、不想装新软件、不想花钱的人。

短板:纯视觉能力弱(看图识物不如 Claude);复杂硬核编程不如第一梯队;微信里体验到的可能还不是全量 Hy3。

成本:API 输入 ¥1/百万 tokens、输出 ¥4/百万 tokens,Apache 2.0 开源。中小团队一天不到 1 块钱。

图片

二、百度搭子:最懂"打工人日常 + 自媒体人"的 Agent

7 月 10 号成都百度 AI Day,百度搭子一口气丢了四个大招:个人版升级、行业首个自媒体专业套件、企业版、搭子联盟。

最值得打工人记住的有三样:

  • 浏览器调用升级:搭子现在能像人一样进网页,点击、跳转、检索、整理信息。你做竞品调研、资料整理时,能看到它一步步操作的过程。
  • 智能路由:任务要高算力就走云端沙箱,要数据安全就走本地沙箱,简单对话走 Chat 模式。搭子自己判断,你不用选。结果:平均任务耗时降了 20%,Token 利用率提升 25%。
  • 行业首个自媒体专业套件:从选题规划、热点洞察,到图文创作、短视频脚本、分镜口播、发布辅助、运营复盘,全在一个对话流里完成。

还有"一镜"数字人制作、"灵医"报告解读等 Skill 上架,企业版支持团队协作与权限管理,搭子联盟启动(中国联通等已加入)。

上线三个月,百度搭子日均提问量增长 20 倍——这个数字比任何参数都实在,说明真有人在拿它干活。

适合谁:做公众号/短视频的新媒体人、要跨应用跨文件干复杂活的人、不想自己切工具的人。

短板:偏"流程套件"和办公协作,深度 coding、技术自建不是它的主战场。

图片

三、美团 LongCat-2.0:给技术团队的"国产算力省钱王"

7 月 6 号,美团把万亿参数大模型 LongCat-2.0 正式开源。

几个数字:总参数 1.6 万亿,平均激活约 480 亿,原生支持 100 万 token 超长上下文。它是业界首个在五万卡国产算力集群上完成全流程训练与推理的万亿参数模型。

重点不是"大",是"国产卡也能跑":

  • 开源同步提供 BF16、FP8、INT8 多精度版本,以及针对国产算力极致优化的推理代码。
  • 意思是:你手里那些存量国产算力卡,不用买新卡、不依赖进口芯片,也能跑起一个万亿模型。
  • 专为 Agentic Coding(智能体编程) 设计。在 SWE-bench Pro 拿了 59.5 分,超过了 GPT-5.5(58.6)和 Claude Opus 4.6(57.3)。

适合谁:技术团队、要自己部署、想盘活国产算力存量卡、在乎数据不出域的企业。

短板:更偏开发者和工程部署,普通打工人用不到;部署有门槛,不是"打开就用"。

图片

一张表,直接抄作业

维度 腾讯 Hy3 百度搭子 美团 LongCat-2.0
一句话定位 微信里的免费 AI 员工 全链路办公 / 自媒体 Agent 国产算力上的开源编程巨兽
最适合 所有人 / 免费党 自媒体人 / 跨应用办公 技术团队 / 自己部署
怎么上手 微信小程序"元宝" 百度搭子 App / 网页 下载权重 + 推理代码部署
要钱吗 元宝 Agent 免费 / API 超便宜 个人版可用 / 企业版付费 开源免费 / 需自有算力
最擅长 做 PPT / Excel / 微信生态 自媒体全链路 / 浏览器操作 Agentic Coding / 长代码
明显短板 纯视觉弱 偏流程套件 部署门槛高

分场景,对号入座

  • "我只会微信,要免费" → 腾讯 Hy3(元宝),打开就能用。
  • "我做公众号/短视频,要一条龙" → 百度搭子(自媒体套件),选题到复盘一个对话流。
  • "我是程序员/技术负责人,要自己部署省算力" → 美团 LongCat-2.0,开源 + 国产卡优化。
  • "我是中小团队老板,要控制成本" → Hy3 API(一天不到 1 块),或 LongCat 自部署盘活存量卡。

算笔账:国产这波,省在哪?

三家逻辑其实一致——降本。

  • Hy3:小团队一天 API 成本不到 1 块钱,元宝 Agent 全免费。
  • LongCat-2.0:让存量国产卡跑万亿模型,不买新卡、不依赖进口芯片。
  • 百度搭子:智能路由把 Token 利用率提了 25%,同样的活更省。

对比海外闭源按次、按量收费,国产在"打工人用得起"这件事上,确实更友好。

避坑:别被参数忽悠,也别乱接数据

  1. 参数 ≠ 你能用上的能力。1.6 万亿不等于你用得上,看"激活参数"和你的实际场景。
  2. 查工具版本。7 月 8 号工信部发了 Claude Code 后门风险提示(2.1.91–2.1.196 版本会回传地域、身份等敏感信息)。用 AI 编程工具前,先升级到安全版本,管好外联权限。
  3. 企业数据别乱接。国产工具在合规上更让人放心,但涉及核心数据的活,先确认部署方式和权限边界。
  4. 先把"我要它干啥"想清楚,再选工具——没有最好的模型,只有最适合你那件事的。

写在最后

这周三家国产大模型,不约而同把口号从"我多强",换成了"我帮你干活"。

对打工人来说,这是好事。AI 终于从秀肌肉,变成了真搬砖。

关注「AI 办公智元」,我会持续追踪这三款国产 AI 在真实办公场景里的表现,下周出每款的深度实测,不吹不黑,只讲真实体验。

你最想让我先测哪一款?评论区告诉我。

 

This article is sourced from: https://mp.weixin.qq.com/s/e4YPJ61y7qGq6hxZ5dBVgw If there are any copyright issues, please contact me at 13951431913#139.com or submit a comment on the right.

The domestic AI scene had a bit of an unusual week.

Baidu, Meituan, and Tencent all unveiled functional, "work-ready" large language models in the same week. These weren't just another chat robot for show; they were genuine announcements that these models are ready to do the heavy lifting.

However, the working population has long since stopped getting excited just by reading parameter sheets. We care about one thing:

Who can actually take over my tasks, and not break the bank doing it?

Today, I'm putting these three "domestic AI employees" side by side for an honest comparison, leaving out the fluff and jumping straight to the conclusion.

图片

I. Tencent Hy3: The Free AI Worker You Can Use Right in WeChat

Released on July 6th, Hunyuan Hy3 is the easiest to jump into out of the bunch.

In plain terms, it's a model designed specifically for executing tasks — not for chatting or writing poetry, but for handling real office work: creating PowerPoint decks, analyzing Excel sheets, generating web pages, and answering questions from the knowledge base. While it boasts a total parameter count of 295 billion, it typically activates only about 21 billion (roughly 7%) when performing tasks.

The real killer feature isn't the parameter count, it's the access point.

Yuanbao (WeChat Mini Program): After integrating Hy3, it gained an Agent function. You can tell it, "Create a report for me on the industry, gather the latest data, and make a PPT," and it will independently search the web, verify the data, run the analysis, and hand you a ready-made PowerPoint file. Best of all, this is completely free.

WorkBuddy (inside Enterprise WeChat): With Hy3, its task success rate jumped from 72% to 90%, and the average processing time was cut by a third. This includes handling meeting minutes, cleaning data, and drawing charts, all within your existing workflow.

WeChat Ecosystem: The public account's AI identity recognition has reached 98.94% accuracy and has quietly been rolled out to over 1 billion users.

Who is this for: People who want to get things done by simply sending a message in WeChat, avoiding the hassle of installing new software or incurring costs.

The downside: Its visual abilities are weak (it can't recognize objects in images as well as Claude), and it isn't the go-to for complex, hardcore programming tasks compared to the top tier. Also, the experience in WeChat may not yet reflect the full power of Hy3.

Cost: The API is priced at ¥1 per million tokens for input and ¥4 per million tokens for output, and it is open-source under Apache 2.0. For a small team, that translates to less than one yuan (roughly 13 cents USD) per day.

图片

II. Baidu Duza: The Agent That Truly Understands "Office Life + Content Creators"

At the Baidu AI Day in Chengdu on July 10th, Baidu Duza dropped four major updates: a personal version upgrade, an industry-first professional kit for content creators, an enterprise version, and the Duza Alliance.

Here are the three most critical points for the average worker:

Browser Interaction Upgrade: Duza can now interact with web pages like a human, clicking, navigating, searching, and organizing information. When you're doing competitor research or compiling materials, you can actually see it stepping through the process.

Intelligent Routing: If a task requires high computational power, it switches to the cloud sandbox; if data security is the priority, it uses the local sandbox; for simple conversations, it defaults to the chat mode. Duza automatically decides this for you. The result is a 20% decrease in average task completion time and a 25% improvement in token utilization.

Industry-First Content Creator Suite: This covers the entire workflow — from topic selection and trend analysis to writing articles, producing short video scripts, handling voiceovers, assisting with publication, and analyzing results — all within a single conversation thread.

Baidu also announced "Yijing" for digital avatar production and "Lingyi" for medical report interpretation, expanded enterprise features for team collaboration and permission management, and the launch of the Duza Alliance (which includes China Unicom).

In just three months since its launch, the number of daily questions asked to Baidu Duza has skyrocketed by 20 times — a statistic far more telling of its usefulness than any parameter figure. This shows that people are genuinely using it to get their work done.

Who is this for: Public account and short video content creators who need a one-stop shop for production, as well as individuals who must work across multiple applications and files but don't want to switch tools manually.

The downside: It leans heavily into process automation and collaboration rather than deep, technical programming or self-built solutions.

图片

III. Meituan LongCat-2.0: The "Domestic Hardware Saver" for Tech Teams

On July 6th, Meituan took the wraps off LongCat-2.0 and made it open-source.

The stats: 1.6 trillion total parameters, an average activation of roughly 48 billion, and native support for an ultra-long context of 1 million tokens. Crucially, it is the first trillion-parameter model in the industry to complete its entire training and inference cycle on a domestically produced, 50,000-core computing cluster.

The focus here isn't that it's large, but that it runs on domestic hardware:

The open-source package includes multiple precision versions (BF16, FP8, INT8) along with inference code optimized specifically for domestic hardware.

What this means is you can run a trillion-parameter model using your existing, domestically sourced GPUs without buying new equipment or relying on imported chips.

LongCat-2.0 is specifically tuned for Agentic Coding (intelligent programming). In the SWE-bench Pro benchmark, it scored 59.5, beating out GPT-5.5 (58.6) and Claude Opus 4.6 (57.3).

Who is this for: Technology teams, companies who need to deploy in-house, businesses looking to maximize their investment in domestic hardware (rather than relying on Nvidia imports), and those prioritizing that their data stays within their environment.

The downside: This is a developer-focused tool for engineering deployment; it isn't something the average office worker would open up for everyday tasks.

图片

一张表,直接抄作业

分场景,对号入座

"I only use WeChat, and it has to be free" → Go for Tencent Hy3 (Yuanbao), which is ready to use right away.

"I run a public account/short video channel and need an all-in-one solution" → Choose Baidu Duza (with the Content Creator Suite) for a seamless flow from idea to publication.

"I'm a programmer/tech manager who needs to deploy in-house and save on compute costs" → Select Meituan LongCat-2.0, which is open-source and optimized for domestic hardware.

"I'm a small business owner focused on cost control" → Consider the Hy3 API (less than 1 yuan a day) or self-deploying LongCat-2.0 to utilize existing hardware.

Breaking it down financially: How do these domestic options save money?

All three companies share a central goal: cost reduction.

Hy3: Small teams can use the API for less than 1 yuan per day, and the Yuanbao Agent is completely free.

LongCat-2.0: Enables the use of existing, domestically built GPUs for trillion-parameter work, eliminating the need to purchase new, expensive foreign hardware.

Baidu Duza: Intelligent routing improves token utilization by 25%, meaning you get more done with fewer computational resources.

Compared to the per-request or per-usage charges of major overseas companies, these domestic solutions are significantly more affordable for the average worker.

Caveats: Don't be duped by parameters, and mind your data security.

Total parameters ≠ usable capability. A 1.6 trillion parameter model might not be relevant if the specific task only uses a fraction of it. Look at "active parameters" and how well the model fits your specific use case.

Check the tool versions. A recent government warning highlighted that specific versions of Claude Code (2.1.91–2.1.196) contain a backdoor that leaks sensitive data. Before using any AI programming tool, ensure you are on a secure version and that external data connections are managed properly.

Don't feed core data into untrusted systems. While domestic tools are generally safer regarding local laws, always verify the deployment model and access permissions before handing over sensitive business information.

Start by defining what you need the AI to do, then choose the tool — there is no single "best" model, only the best model for your specific job.

写在最后

This week, these three domestic players shifted their messaging from "how powerful I am" to "how I can help you work."

For the workforce, this is a positive development. AI is transitioning from a tech show-off to a productive co-worker.

I will continue tracking these three tools as they are applied to real-world office tasks. Stay tuned for full, no-nonsense reviews of their performance in the coming weeks.

Which one should I review first? Let me know in the comments.

暂无评论,快来发表第一条评论吧!No comments yet. Be the first!

📮 需求咨询📮 Inquiry