本文来源:https://mp.weixin.qq.com/s/ypEqcKe78B8N3BXcEIbmfg
若有版权问题,请来信告知:13951431913#139.com 或在右方提交评论。
干活必须上最好的模型?
四大国产模型同题竞技
谁合适,谁就上
PPT · 信息图 · 公众号排版 · 长文润色 · 四轮实测
百度千帆 Token Plan · 实测
OPENCODESKILL大家好啊,我是等待 WAIC 开幕的甲木,
人已经到上海两天了,准备迎接 WAIC 开幕,这次规模非常大,大会的票前两周就被抢光了……这次依旧我们去年的传统,会给大家带来「机器人的探展直播」。
再来回答一下大家关心的问题,前段时间不是开源了「公众号排版 Skill」,三天 2.1k Star,大家的反馈还是可以的。
但是也有很多朋友好奇,国内的模型跑出来效果怎么样,以及最近模型更新很快。今天这个刷榜,明天那个升级。
可真轮到自己干活,真的有必要上最好的模型吗?
为了验证这个问题,我把我最近主要的几个需求,用百度千帆 Token Plan 里的 DeepSeek-V4-Pro、GLM-5.2、Kimi-K2.6 和 ERNIE 5.1 轮番跑了PPT、HTML 信息图、公众号排版和写作四轮任务。
再用 Codex 进行评估。
这一跑就是一整晚,还把我Codex 的周限额也烧完了。还好今天早上额度重置,不然这篇可能真要卡在这里了 = =
先来说结论:
DeepSeek
数据密集、需要把逻辑压紧的任务,可以先试
GLM
强调结构和边界的任务,可以试
Kimi
长文和视觉表达,可以试
ERNIE
中文语境下的表达交付,也可以试
对了,简单介绍一下百度千帆 Token Plan,它是一个面向个人开发者的多模型订阅计划。
以我这次用的 Pro 档为例,200 元一个月,包含 2.3 亿 Tokens。多款主流国产模型,共用这一份额度。
— 支持的主流模型
现在模型能力上来了,再叠加Skill 的硬约束,很多模型其实都能做得不错。
好,我们直接看实战场景,拿结果说话。
以下几个场景为了方便,我直接用 OpenCode Agent 叠加百度千帆 Token Plan 计划来跑,多模型切换极其方便。
📦 4 Parts + Conclusion
👉 滑动
PART 01
PPT 实测
同题四种讲法
PART 02
HTML 图
风格统一个性在
PART 03
公众号排版
差距缩得很快
PART 04
公众号写作
比事实比修改量
PART ///
写在最后
选模型只看三点
01
PART
PPT:同一份材料,四种讲法
PPT BATTLE · ROUND 01
第一轮,测 PPT。
藏师傅的 PPT Skill 爆火,而且PPT 永远是打工人逃不过的槛,
譬如我现在要给老板做汇报,我会直接调用guizang-pptSkill,让它生成一份可以继续修改的 html。既方便演示,又美观。
我直接输入同一份 prompt,
— OpenCode 中同一份任务的 prompt
然后接入百度千帆里面的不同模型,我们来看看四个模型生成的封面和数据页。
— OpenCode 模型接入百度千帆 Token Plan
先看封面:
— PPT 封面四种模型效果
再看数据页:
乍一看都挺像那么回事,排版也符合 Skill 规范。但仔细看,你会发现它们讲故事的习惯很不一样。
DeepSeek-V4-Pro 的页面秩序更规整,标题、层级和节奏都稳。
Kimi-K2.6 和 GLM-5.2 很相似,都喜欢把信息压紧,数字和结论挨得很近,比较适合业务分析和复盘。
ERNIE 5.1 则绘制了一个矩形图。该有的数据页、流程页和结论页都在,还挺有设计感。
如果只问谁“会不会做 PPT”,这轮分不出高下。
02
PART
HTML 信息图:风格被统一了,个性还在
SOCIAL CARD · ROUND 02
第二轮再来试试藏师傅的社交信息卡片的效果,小红书、小绿书的贴图神器,
直接调用guizang-social-cardSkill。
要求是生成 6 张 HTML 信息卡,再做一组公众号封面。所有页面统一使用Swiss Style 和 IKB 克莱因蓝,不联网找图,也不用 AI 生图,只靠文字和排版完成。
— OpenCode 中同一份任务的 prompt
来看看生成的效果:
— 信息图封面
— 信息图检查页
这一轮最直观的感受,是——
Skill 的强约束真的能把风格拉到同一条线上
网格、配色、标题层级都被 Skill 固定住了,模型没有太多乱发挥的空间。差异留在更细的地方:有人偏几何秩序,有人更在意阅读路径,也有人会把标题和组件做得更饱满。
跟上轮测试一样,Kimi-K2.6 很喜欢做对比,ERNIE 5.1 则又开始绘制图表。
这其实很像给四位设计师发了同一套瑞幸咖啡的品牌手册。最后的作品可能各有千秋,但都遵循了一样的品牌风格,让人看的出来是什么牌子。
03
PART
公众号排版:差距缩得比我想象中快
GZH DESIGN · ROUND 03
第三轮,就是我跟小李一起合作开源的公众号排版 Skill,这个 Skill 是把同一篇文章交给四个模型。
标题区、引用、分节、重点标注和总结组件都要做,最后还得输出一份可以复制进公众号继续编辑的 HTML。
这里其实不用这么复杂的 prompt,直接说一句用摸鱼绿风格帮我把XX文章进行排版就可以了……
先看 DeepSeek 和 GLM:
再看 Kimi 和 ERNIE:
乍一看,基本看不出任何差异是不是?
拿掉模型名,让我只看成品,我未必能一眼猜出分别是谁做的,都很美观。
字号、间距、组件、兼容性和复制效果,过去都要靠模型临场发挥。
现在 Skill 把它们写成了规则,模型只需要在边界内组织内容。
说白了,Skill 先把最容易翻车的地方管住了,这个时候模型之间的差异就很小了。
所以,gzh-design还是基本适配国内所有模型的,欢迎大家使用!
04
PART
公众号写作:别截金句,要看同一段怎么改
HUMANIZER · ROUND 04
之前很多人说 AI 生成的内容 AI 味儿比较浓,
这里我们来试试不同模型的语言表述,不比辞藻,比事实和修改量。
最后一轮,我用Humanizer-zhSkill 做公众号长文润色。
原稿还是那份 6 周工作流案例,有固定事实、固定数字,也写清了能力边界。
四个模型都要在不改事实的前提下,把空话、套话和 AI 味清掉。
再看看能力边界段,
在写作方面上,就可以看出不同模型的区别。
DeepSeek 会主动加判断,句子更硬
GLM 会把因果和边界解释清楚,收束得比较稳
Kimi 读起来顺,面向普通读者时比较亲和
ERNIE 倾向把事情交代完整,结尾会落到具体动作上
但是四个模型都有不同程度的 AI 味,
写作这件事,模型们属于是越发展越退步了……
很长一段时间还是得依赖人类的品味和判断。
///
LAST
写在最后
CONCLUSION · HOW TO CHOOSE
其实大家也能从这四个任务里面看出来,在涉及到这种编码场景的情况下,所有的 AI 模型当前都完成得差不多。
我们会发现,整个国产模型当前完成的一些效果也都是比较良好的。
在 Skill 有强约束的情况下,很多模型都可以很好的交付结果
没必要用高射炮打蚊子
它们当然还有差异。只是这种差异越来越像分工,而不是“一台能干活,另外三台不行”。
我现在选模型,只看三点:
能力
模型的能力边界
价格
是否跟能力匹配
交付
能否稳定按 Skill 交付
这也是 TokenPlan 的实际价值。
同一个计划里能切换多种模型,写作、排版、PPT 和视觉任务不用绑死在一家模型上,谁适合当前任务,就让谁上。
顺带一提,百度千帆TokenPlan还挺稳的。这个批量测试我从晚上 8 点开始,整晚没有遇到中断。跑完这一大轮,后台用量大约消耗了 8%。
— 非常耐用!
我买的,不只是更多 Token。
是需要干活时,能从同一个模型池里挑到合适的那个,少排队、少重试,别让工具把自己的工作心流打断。
Source: https://mp.weixin.qq.com/s/ypEqcKe78B8N3BXcEIbmfg If there are copyright issues, please notify us via email: 13951431913#139.com or submit a comment on the right.
Do you really need the best model to get work done?
Four major domestic models compete on the same tasks
Whoever fits, goes in
PPT · Infographic · WeChat article layout · Long-form polishing · Four rounds of real-world testing
Baidu Qianfan Token Plan · Real-world test
Hi everyone, I'm Jiamu, waiting for WAIC to open.
I've been in Shanghai for two days now, getting ready for the WAIC opening. This year's event is massive — tickets were completely sold out two weeks ago... Following our tradition from last year, we'll be bringing you a "robot exhibition livestream."
Let me also address a question many of you have asked. Earlier, I open-sourced the "WeChat article layout Skill," which hit 2.1k stars in three days. The feedback has been pretty good.
But many friends are also curious about how well domestic models perform, especially since models are updating so fast. Today one tops the leaderboard, tomorrow another gets an upgrade.
But when it's actually time to get work done, is it really necessary to use the best model?
To test this, I took my main recent needs and ran them through DeepSeek-V4-Pro, GLM-5.2, Kimi-K2.6, and ERNIE 5.1 in the Baidu Qianfan Token Plan across four rounds of tasks: PPT, HTML infographic, WeChat article layout, and writing.
Then I used Codex to evaluate the results.
The whole run took all night and burned through my Codex weekly quota. Luckily, the quota reset this morning — otherwise this article might have gotten stuck right there = =
First, the conclusions:
DeepSeek
For data-heavy tasks that need tight logic, try it first
GLM
For tasks that emphasize structure and boundaries, give it a try
Kimi
For long-form writing and visual expression, give it a try
ERNIE
For expression and delivery in Chinese contexts, give it a try too
By the way, let me briefly introduce the Baidu Qianfan Token Plan. It's a multi-model subscription plan for individual developers.
Take the Pro tier I used this time: 200 yuan per month, including 230 million tokens. Multiple mainstream domestic models share this quota.
— Supported mainstream models
Now that model capabilities have improved, combined with the hard constraints of Skills, many models can actually do quite well.
Okay, let's go straight to the real-world scenarios and let the results speak.
For convenience, I ran the following scenarios directly using OpenCode Agent with the Baidu Qianfan Token Plan — switching between models is extremely easy.
📦 4 Parts + Conclusion
👉 Swipe
PART 01
PPT Real-world test
Four approaches to the same topic
PART 02
HTML graphics
Unified style, personality intact
PART 03
WeChat article layout
The gap is closing fast
PART 04
WeChat article writing
Comparing facts and amount of revision
PART ///
Final thoughts
Only three criteria for choosing a model
01
PART
PPT: The same material, four different approaches
PPT BATTLE · ROUND 01
Round one: testing PPT.
Zang's PPT Skill went viral, and PPTs are forever an unavoidable hurdle for office workers.
For example, when I need to give a report to my boss, I directly call guizang-pptSkill and have it generate an HTML file that I can continue editing. It's convenient for presentations and looks great.
I entered the same prompt directly.
— The same task prompt in OpenCode
Then I connected the different models in Baidu Qianfan and looked at the cover pages and data pages generated by the four models.
— OpenCode models connected to Baidu Qianfan Token Plan
First, the cover pages:
— PPT covers from four models
Then the data pages:
At first glance, they all look the part, and the layouts follow the Skill's specifications. But looking closely, you'll notice their storytelling habits are quite different.
DeepSeek-V4-Pro's pages are more orderly — titles, hierarchy, and rhythm are all steady.
Kimi-K2.6 and GLM-5.2 are very similar; both like to compress information, with numbers and conclusions sitting close together, which suits business analysis and retrospectives.
ERNIE 5.1, meanwhile, drew a rectangular chart. It has all the expected data pages, process pages, and conclusion pages, and it has a nice sense of design.
If we only ask who "can make a PPT," this round produces no clear winner.
02
PART
HTML infographic: The style is unified, but the personality remains
SOCIAL CARD · ROUND 02
In round two, we test Zang's social info card effect — the go-to tool for Xiaohongshu and Little Green Book image posts.
I directly call the guizang-social-card Skill.
The requirement is to generate 6 HTML info cards plus a set of WeChat article covers. All pages must use Swiss Style and IKB Klein Blue, with no online image searching and no AI image generation — only text and layout.
— The same task prompt in OpenCode
Here are the results:
— Infographic covers
— Infographic inspection page
The most immediate takeaway this round:
The Skill's strong constraints really do pull the style onto the same line
Grid, color scheme, and title hierarchy are all locked down by the Skill, leaving the models little room to improvise. The differences remain in the finer details: some lean toward geometric order, some care more about reading paths, and some make titles and components fuller.
Same as the last round, Kimi-K2.6 loves making comparisons, while ERNIE 5.1 starts drawing charts again.
This is a lot like handing four designers the same Luckin Coffee brand guideline. The final pieces may each have their own merits, but they all follow the same brand style, making it clear which brand they belong to.
03
PART
WeChat article layout: The gap closed faster than I expected
GZH DESIGN · ROUND 03
In round three, we use the WeChat article layout Skill that I co-developed with Xiao Li. This Skill sends the same article to all four models.
Each one must produce the title area, blockquotes, sections, key highlights, and summary components, and finally output an HTML file that can be copied into WeChat for further editing.
You don't actually need such a complex prompt here — you can just say, "Please format this article for me in the moyu-green style."
First, DeepSeek and GLM:
Then Kimi and ERNIE:
At first glance, you basically can't see any difference, right?
If you removed the model names and showed me only the finished pieces, I probably couldn't tell at a glance who made which. They all look great.
Font sizes, spacing, components, compatibility, and copy-paste behavior all used to depend on the model's on-the-spot performance.
Now the Skill has turned them into rules — the model only needs to organize content within the boundaries.
In other words, the Skill manages the parts most likely to go wrong, so the differences between models become very small.
So gzh-design is basically compatible with all domestic models. You're welcome to use it!
04
PART
WeChat article writing: Don't quote the highlights — look at how the same passage gets revised
HUMANIZER · ROUND 04
Many people have said that AI-generated content has a strong "AI flavor."
So here we test the language expression of different models — not comparing rhetoric, but comparing facts and the amount of revision.
In the final round, I used the Humanizer-zh Skill to polish a long WeChat article.
The source draft is the same 6-week workflow case study, with fixed facts and fixed numbers, and it clearly states the capability boundaries.
All four models had to remove empty talk, clichés, and AI flavor without changing any facts.
Now look at the capability boundary section:
In terms of writing, you can clearly see the differences between the models.
DeepSeek proactively adds judgments — the sentences are tougher
GLM explains causality and boundaries clearly and wraps up steadily
Kimi reads smoothly and feels more approachable for ordinary readers
ERNIE tends to explain things thoroughly and ends with concrete actions
But all four models still have some degree of AI flavor.
When it comes to writing, the models seem to be regressing as they develop...
For a long time to come, we'll still have to rely on human taste and judgment.
///
LAST
Final thoughts
CONCLUSION · HOW TO CHOOSE
As you can probably tell from these four tasks, when it comes to coding-related scenarios, all AI models currently perform more or less the same.
We can see that domestic models are generally producing quite good results across the board.
With strong constraints from Skills, many models can deliver great results
No need to use a cannon to kill a mosquito
Of course, there are still differences. But those differences increasingly look like a division of labor, not "one works and the other three don't."
When I choose a model now, I only look at three things:
Capability
The model's capability boundaries
Price
Whether it matches the capability
Delivery
Whether it can reliably deliver according to the Skill
This is the real value of the Token Plan.
You can switch between multiple models within one plan. Writing, layout, PPT, and visual tasks don't have to be locked to a single model — whoever fits the current task gets to go in.
By the way, the Baidu Qianfan Token Plan is pretty stable. I started this batch test at 8 PM and didn't hit a single interruption all night. After running this whole big round, the backend usage consumed about 8% of the quota.
— Very durable!
What I bought isn't just more tokens.
It's the ability to pick the right model from the same pool when I need to get work done — fewer queues, fewer retries, and no tools breaking my workflow.