洛洛的第18篇原创文章
大家好,我是洛洛。
最近大模型那么多,到底哪个更适合普通白领呢?
今天我们就拿白领最常遇到的读文档、核对数据、做活动页面这些场景来测试一下。
这些场景大模型表现如何,光看跑分根本选不出来。我们直接拿我一个做营销的朋友的真活儿上手,把 MiniMax M3、DeepSeek V4 Pro、Kimi 2.7、Gemini 3.5 Flash 摆到一起各跑一遍。
看看这种典型的办公室白领的工作,哪个模型更能打?
案例实测
朋友有个咖啡品牌叫山野咖啡,主打云南保山产地豆和冷萃工艺,走的是产地故事加质价比路线,价位卡在瑞幸和星巴克中间。最近她要做一个活动营销方案,并且需要看到活动落地页的效果,我们一步一步走完这个流程。
第一步:读参考资料出方案
朋友给我了这个项目的参考资料,2 张友商活动页长截图,另外还有一份15万字188页的山野咖啡产品文档及用户访谈摘录(脱敏)。
这些资料靠人一条条看完再提炼,半天时间就没了。我想看看模型能不能读懂复杂的长图和这15万字的文档,还能根据这些资料,直接给出一版能用的活动页方案。
我直接在4个模型中都输入同样的内容:
DeepSeek V4 Pro
DeepSeek可以同时处理截图和长文档,产出的方案也精准兼顾了高级调性与硬核卖点,用低饱和视觉与情绪文案维持品牌审美,用18.8元定价、0糖0脂、科学工艺及空瓶回收玩法,直击用户痛点并巧妙促成高裂变与高复购。
DeepSeek可以坦率地承认自己有读不出来的地方,主要是一些小字,细则。整体来说表现还是不错的。
Kimi 2.7
Kimi 长文本是老本行,在长文本阅读上还是非常强的。整个方案基于50条真实用户访谈,避开瑞幸低价与星巴克社交,以产地情绪加慢工艺错位竞争,用硬核数据建立信任,让促销不损调性。
对于读不清楚的地方,Kimi 2.7很诚实,写得也相对比较清楚。读不清的地方和DeepSeek差不多,主要是一些小字。
MiniMax M3
M3是唯一一个给我输出了完整文档供我下载的,它这个方案的完整程度,是可以直接交给老板看的程度了。方案内容也不错,差异化清晰,用的产地心智战。每条决策都有访谈原话兜底,SKU 只推 3 款很克制,是真的懂营销。
关于长截图,MiniMax M3很诚实地告诉我它只能读清80%,哪些是读不清楚的,什么原因造成的,都列得清清楚楚。其实我不怕大模型读不清,就怕它读不清还瞎编。这一点MiniMax M3确实做得不错。
Gemini 3.5 Flash
Gemini 3.5 Flash 主打快和多模态,作为海外参照,图文理解的综合精度、中文文档和咖啡用户反馈把握得也比较准确,速度也确实更快。方案精准兼顾了高级调性与硬核卖点,整体表现还是不错的。
Gemini 3.5 Flash对于对不清的地方就写得相对比较笼统了。也没有一个大致的能读清的比重,我就不好一下子非常直观地知道它到底懂读了多少。
读1M长文本和识别长截图再生成活动方案这个步骤,4个模型整体变现都不错。DeepSeek V4 Pro能完成任务,整体表现中规中矩,Gemini 3.5 Flash胜在生成的速度更快,Kimi 2.7对长文本的把握会更准确一些,MiniMax M3生成的方案更完整可以直接交付,对读不清楚的截图内容的描述也更加准确和细致。
第二步:核实活动页面核心数据
方案里准备在页面引一句给品牌背书的市场数据,草稿写的是中国咖啡市场即将破万亿。我需要核对这个说法到底准不准确。
数据不对,活动页轻则被同行挑刺,重则涉嫌虚假宣传。
提示词:
我的活动页方案里准备引一句中国咖啡市场即将突破万亿。请帮我核对,第一,万亿口径说的是整个咖啡行业还是现磨咖啡,最新可查的市场规模数字大概是多少、来源和年份是什么;第二,作为体量参照,拉一下星巴克最近一个财年的营收和当前市值;第三,给页面一句更严谨、不会被挑刺的表述建议。
DeepSeek V4 Pro
DeepSeek虽然没有接入专门的行业数据库,但是给出的答案还是靠谱的,而且还很贴心地给出了三种表述建议。
Kimi 2.7
Kimi查到了“万亿”口径的来源,也指出这种表达可能面临的风险,专业性还是不错的。
MiniMax M3
MiniMax 网页端推出金融模块,我们这个case正好可以试一下。
在对话框点击“+”,选择已连接插件,再选择金融数据库即可。
走 Agent 金融模块能直接连库取实时行情,星巴克的美股营收和市值能即时拉到,数据时效是最大优势。
MiniMax M3是唯一一个帮我验证了数据信源的,这一点确实让我心里更踏实一些。
Gemini 3.5 Flash
Gemini 3.5 Flash 在这个问题上的生成速度同样是最快的,生成的结果也可以用。
这一轮从给出的答案是否正确上,各家模型拉不开差距,从生生速度上Gemini 3.5 Flash更快,更占优势。不过从信源可信度上,MiniMax M3会更好一些。
第三步:把方案和核准的数据变成落地页
这一步我目标就是让各个模型自由发挥。
提示词:
根据前面定下的「山野冷萃」活动页方案,用单个 HTML 文件做一版活动落地页。
DeepSeek V4 Pro
看一下DeepSeek V4 Pro自由发挥的结果,只能说中规中矩,这个水平的前端页面,半年之前的模型其实也可以达到。
Kimi 2.7
Kimi 2.7生成的速度是真慢啊,不过这个页面的效果很好,不管是配色还是交互都非常高级,按钮和文字排版这些细节也处理得很到位,是可以直接拿去上线的程度。
MiniMax M3
这一轮M3的生成速度也比较慢,但是整体效果很不错,山野自然风的版式很好看、还有3D效果、领券按钮和好评轮播的交互也很顺,整体配色也很有高级感。
Gemini 3.5 Flash
Gemini 3.5 Flash 前端出活儿快,不过第一版打开竟然还是代码。
好不容易打开页面之后发现,还不如不打开。这根本不是一个可以落地的活动页。
各家模型真正拉开差距的是在落地页生成这一步,出页面比的是审美加完成度,能少返工、不用再补配色的产出最省心。Kimi 2.7和 MiniMax M3 在国产里靠前,这两个都是能直接上线的水准。
洛洛有话说
三个 case 跑完,我还挺意外的。读1M长文本、认长图、核对数据,几家基本拉不开差距,能力都到位了。可一旦放开手让几家自由发挥生成活动落地页,差距一下就拉开了。
DeepSeek V4 Pro 像稳定发挥的中等生,该有的都有,少了点让人眼前一亮的地方。Kimi 2.7 的网页生成是真的惊喜,高级感很强,出来就能直接用。MiniMax M3 同样可圈可点,核数据时连着金融数据库,几家结论虽然一致, M3 还附上了可靠信源,可信度一下就上去了,前端的 3D 视差和滚动动效也做得到位。Kimi 和 M3 都偏慢,但慢得值,生成的页面基本直接能上线。Gemini 3.5 Flash 速度最快,可惜出的页面一开始打不开,好不容易打开也不像个落地页,快归快,用不上。
四个顶流横评下来,哪个更适合办公室白领?其实还是看各自具体的场景,不过从读文档、核数据、页面生成以及组合着用最省,返工最少的角度,Kimi2.7和MiniMax M3都是不错的选择。
如果你的工作对数据准确性要求特别高的话,MiniMax M3可能会更适合一些。
好咖啡讲究产地溯源,好模型也一样,能交出结果、还交代得清来路的,才真正让人放心。
测评结果各位读者老师体验一致吗?评论区聊聊~
往期推荐:
AI 看得越多越焦虑?6 本书 + 3 门课,帮你把认知一次补齐
马斯克、哈萨比斯都绕不开的15部AI影片,看完聊AI不再外行(建议收藏)
Luo Luo's 18th original article
Hello everyone, I'm Luo Luo.
With so many large models available these days, which one is actually better suited for ordinary office workers?
Today we'll test this using scenarios that office workers commonly encounter: reading documents, verifying data, and creating activity landing pages.
You can't pick the best model just by looking at benchmark scores. We put MiniMax M3, DeepSeek V4 Pro, Kimi 2.7, and Gemini 3.5 Flash to the test using real work from my marketing friend.
Let's see which model handles typical office work the best.
Case Study
My friend has a coffee brand called Shanye Coffee, which focuses on Yunnan Baoshan origin beans and cold brew craftsmanship. It positions itself as origin-story driven with quality pricing, sitting between Luckin and Starbucks in price. Recently she needed an activity marketing plan and a landing page mockup, so we walked through the whole process step by step.
Step One: Read Reference Materials and Draft a Plan
My friend sent me the project materials: two long screenshots of competitor activity pages, plus a 150,000-word, 188-page Shanye Coffee product document and anonymized user interview excerpts.
Going through these documents line by line would take half a day. I wanted to see whether the models could understand complex long screenshots and a 150K-word document, and then directly produce a usable landing page plan based on them.
I fed the same content into all four models:
DeepSeek V4 Pro
DeepSeek can simultaneously process screenshots and long documents. Its output plan accurately balances premium branding and hardcore selling points, using low-saturation visuals and emotional copy to maintain brand aesthetics, while its 18.8 RMB pricing, zero-sugar zero-fat claim, scientific craftsmanship, and empty-bottle recycling mechanics directly hit user pain points and cleverly drive high virality and repeat purchases.
DeepSeek honestly admitted it couldn't read certain parts, mainly small print and fine details. Overall, its performance was quite good.
Kimi 2.7
Long-text processing is Kimi's specialty, and it remains very strong at it. The entire plan is based on 50 real user interviews, positioning itself against Luckin's low pricing and Starbucks's social atmosphere by leveraging origin emotion and slow craftsmanship as a differentiation strategy, using hard-core data to build trust while keeping promotions from damaging brand tone.
For parts it couldn't read clearly, Kimi 2.7 was quite honest and wrote relatively clearly about it. The unreadable parts were similar to DeepSeek's—mainly small text.
MiniMax M3
M3 was the only one to output a complete document I could download. The plan's completeness was at a level ready to be presented directly to a boss. The content was also solid, with clear differentiation and a position-ing strategy built on origin perception. Every decision was backed by original interview quotes, and it kept the SKU to just three products—truly understanding marketing.
Regarding the long screenshots, MiniMax M3 honestly told me it could only read 80% clearly, and it listed exactly which parts were unclear and what caused the issues. I don't mind large models not reading everything—I just hate it when they make things up. MiniMax M3 really did well here.
Gemini 3.5 Flash
Gemini 3.5 Flash excels at speed and multimodal processing. As a foreign reference, its integrated accuracy in image-text understanding, grasp of Chinese documents, and interpretation of coffee user feedback were both accurate, and its speed was indeed faster. The plan accurately balanced premium tone with hardcore selling points, performing well overall.
Gemini 3.5 Flash was relatively vague about what it couldn't read clearly. It also didn't provide a rough percentage of what it could read, making it harder to intuitively gauge how much it actually understood.
For reading 1M-long text, recognizing long screenshots, and generating activity plans, all four models performed well overall. DeepSeek V4 Pro completed the task with solid, mid-range performance. Gemini 3.5 Flash had the faster generation speed. Kimi 2.7 grasped long texts more accurately. MiniMax M3 generated the most complete plan that could be delivered directly, with more accurate and detailed descriptions of unreadable screenshot content.
Step Two: Verify Core Data on the Activity Page
The plan includes a market data quote to endorse the brand on the page—the draft reads "China's coffee market is about to break 1 trillion yuan." I needed to verify whether this claim is accurate.
Incorrect data can cause competitors to pick flaws in minor cases, or worse, raise false advertising concerns.
Prompt:
My activity page plan includes a statement that China's coffee market is about to exceed 1 trillion yuan. Please help me verify: First, does the "1 trillion" figure refer to the entire coffee industry or freshly ground coffee? What is the latest verifiable market size number, its source, and the year? Second, as a volume reference, pull up Starbucks's revenue and current market cap for its most recent fiscal year. Third, suggest a more rigorous phrasing for the page that won't be easily challenged.
DeepSeek V4 Pro
Although DeepSeek doesn't have a dedicated industry database connected, its answers were still reliable, and it thoughtfully provided three phrasing suggestions.
Kimi 2.7
Kimi found the source of the "1 trillion" figure and also pointed out the potential risks of such phrasing. Its professionalism was quite good.
MiniMax M3
MiniMax's web version offers a financial module, which was perfect for testing in our case.
Click "+" in the chat box, select connected plugins, and choose the financial database.
Using the Agent financial module connects directly to the database to pull real-time quotes. Starbucks's US stock revenue and market cap can be fetched instantly, making real-time data timeliness its biggest advantage.
MiniMax M3 was the only one that verified the data sources for me, which gave me more confidence.
Gemini 3.5 Flash
Gemini 3.5 Flash was again the fastest at generating a response here, and its output was usable.
In this round, there wasn't much gap between models on answer correctness. Gemini 3.5 Flash held the advantage in speed. However, on data-source credibility, MiniMax M3 performed better.
Step Three: Turn the Plan and Verified Data into a Landing Page
For this step, my goal was to let each model run freely.
Prompt:
Based on the finalized "Shanye Cold Brew" activity page plan, create a single HTML file to make an activity landing page.
DeepSeek V4 Pro
Looking at DeepSeek V4 Pro's free-form result—only so-so. The frontend page quality at this level could have been achieved by models from six months ago.
Followed
Follow
Replay Share Like
Close
Watch More
More
Exit Full Screen
Luo Luo AI Says Followed
Share Video
, Duration 00:09
0/0
00:00/00:09
Switch to Landscape Mode
Continue Playing
Progress Bar, 0%
Play
00:00
00:09
00:09
Playback Speed
Full Screen
Playing at Playback Speed
0.5x 0.75x 1.0x 1.5x 2.0x
Ultra HD Smooth
Continue Watching
Comparing 4 Top-Tier Models: Which Is More Suitable for Office Workers?
Watch More
Repost
,
Comparing 4 Top-Tier Models: Which Is More Suitable for Office Workers?
Luo Luo AI Says Followed
Share Like Bookmark
Synced to Look Look, write your comment
Video Details
Kimi 2.7
Kimi 2.7 was really slow to generate, but the page result was excellent—both the color scheme and interactions felt very premium. Details like button and text layout were handled well, at a level ready for launch.
Followed
Follow
Replay Share Like
Close
Watch More
More
Exit Full Screen
Luo Luo AI Says Followed
Share Video
, Duration 00:12
0/0
00:00/00:12
Switch to Landscape Mode
Continue Playing
Progress Bar, 0%
Play
00:00
00:12
00:12
Playback Speed
Full Screen
Playing at Playback Speed
0.5x 0.75x 1.0x 1.5x 2.0x
Ultra HD Smooth
Continue Watching
Comparing 4 Top-Tier Models: Which Is More Suitable for Office Workers?
Watch More
Repost
,
Comparing 4 Top-Tier Models: Which Is More Suitable for Office Workers?
Luo Luo AI Says Followed
Share Like Bookmark
Synced to Look Look, write your comment
Video Details
MiniMax M3
M3 was also relatively slow in this round, but the overall effect was great. The Shanye natural-style layout looked beautiful, with 3D effects, smooth coupon-claim button interactions, and a nice review carousel. The overall color palette also felt premium.
Followed
Follow
Replay Share Like
Close
Watch More
More
Exit Full Screen
Luo Luo AI Says Followed
Share Video
, Duration 00:29
0/0
00:00/00:29
Switch to Landscape Mode
Continue Playing
Progress Bar, 0%
Play
00:00
00:29
00:29
Playback Speed
Full Screen
Playing at Playback Speed
0.5x 0.75x 1.0x 1.5x 2.0x
Ultra HD Smooth
Continue Watching
Comparing 4 Top-Tier Models: Which Is More Suitable for Office Workers?
Watch More
Repost
,
Comparing 4 Top-Tier Models: Which Is More Suitable for Office Workers?
Luo Luo AI Says Followed
Share Like Bookmark
Synced to Look Look, write your comment
Video Details
Gemini 3.5 Flash
Gemini 3.5 Flash generated the frontend quickly, but the first version opened as code instead of a rendered page.
After finally getting the page to open, it turned out it would have been better not to open it at all. This was far from a deployable activity page.
The real gap between models showed up at this landing page generation step. Producing a page is about aesthetics plus completeness—the output that requires the least rework and no additional color matching is the easiest to use. Kimi 2.7 and MiniMax M3 lead among domestic models, both at a level ready for launch.
Luo Luo's Thoughts
After running all three cases, I was quite surprised. Reading 1M-long text, recognizing long images, and verifying data—there wasn't much difference between the models, and their capabilities were all up to par. But once I let them free-generate activity landing pages, the gap widened immediately.
DeepSeek V4 Pro is like a steady middle-performing student: it has everything needed, but lacks something truly eye-catching. Kimi 2.7's webpage generation was genuinely surprising—very premium, ready to use right out of the gate. MiniMax M3 was also noteworthy: during data verification, it connected to the financial database. While all models' conclusions aligned, M3 additionally attached reliable sources, boosting trustworthiness. Its frontend's 3D parallax and scroll animations were also well executed. Both Kimi and M3 are on the slower side, but the slowness is worth it—the generated pages are basically launch-ready. Gemini 3.5 Flash was the fastest, but unfortunately its page couldn't open initially, and even when it finally did, it didn't look like a proper landing page. Fast as it is, it's unusable.
After comparing four top-tier models, which is better suited for office workers? It ultimately depends on the specific scenario. But from the angles of reading documents, verifying data, generating pages, and minimizing rework through combined usage, both Kimi 2.7 and MiniMax M3 are solid choices.
If your work demands particularly high data accuracy, MiniMax M3 may be the better fit.
Just as good coffee values origin traceability, good models do too—those that deliver results and also explain where they came from are truly trustworthy.
Did our readers get the same experience with the test results? Share your thoughts in the comments!
Previous Recommendations:
Anxious from watching too much AI content? 6 books + 3 courses to fill your knowledge gaps at once
15 AI films that neither Musk nor Hassabis can avoid—finish watching and stop being an AI novice (Recommended to save)
June latest roundup: Domestic and international token freebie collection (Recommended to save)!
























