In the era of large models, code has indeed become increasingly cheap. Big data pushes a bunch of AI-generated web pages every day, and after seeing too many, it really gets a bit aesthetically tiring.
These web pages always have a strong AI flavor, the same rounded cards, card shadows, Tailwind purple, emoji packs…
Just yesterday, I was flooded by the K2.5 model released by Kimi. This time, what they are highlighting isvisual programming, where a screenshot or a video can replicate an almost identical web page.
Today, let's test together and see whether Kimi K2.5's "eyesight" and "taste" really hold up?
01. Visual Programming Evaluation
First, we directly sent a screenshot of a good-looking template to Kimi.
The prompt was very simple: replicate this website.
It was indeed impressive. You can tell that K2.5 recognizes and extracts the visual assets in the image and turns the image into interactive web page code.
It also automatically made a responsive layout, and you can directly switch between mobile/desktop web preview.
Also, the web pages generated by their Agent have a very user-friendly design: as long as you click the edit button in the upper right corner, you can directly select the place you need to modify in the preview interface, tell Kimi in plain language where to change it, and it can adjust precisely.
For the generated web page, K2.5 will conveniently help us deploy it online. Click Share in the upper right corner of the page, and you can directly copy the access link.
Replication address:https://rqgcia4tyj2lk.beta-ok.kimi.link
Let's further examine its video replication capability.
The prompt continues with the simple style:Help me replicate an identical web page。
Kimi K2.5 captured the visual features in the video quite accurately. The layout, color matching, and information hierarchy were all restored very well, and it looks as if it were made by copying it. The AI flavor is indeed much lighter.
02. Visual Understanding Evaluation
Kimi K2.5's most critical underlying change is its native visual capability,it can directly understand the interface structure and details in images and videos。
So next, let's examine its visual capability in non-programming fields.
This is a photo of Beijing's Yao'er Hutong. Apart from the hutong element, there are no prominent landmarks or text clues in the image. Let's see what Kimi can tell.
Prompt: determine where this is.
Kimi recognized that this is an old hutong in Beijing's Second Ring Road and narrowed down a few approximate locations, but getting it to directly say Yao'er Hutong still seems a bit difficult.
Let's try again and have Kimi estimate the height of the young man near the car door in the image.
Kimi answered almost instantly. It told me that the young man's height is about between 183-188 cm, most likely around 185 cm.
Gong Jun's official height is 186 cm, very close! (I admit, I was somewhat biased on this question)
Let's continue raising the difficulty. I found a real civil service aptitude test question from 2025 and sent it to Kimi.
Prompt: answer this civil service aptitude test question.
After a brief reasoning process, Kimi quickly grasped that the key to this question lies in the relative positions of the same patterns, and thus chose the correct answer D.
At this point, K2.5's visual capability is already very intuitive.Although guessing the location may not allow precise positioning, visual reasoning is a strong suit, and it can extract effective information from the image to support reasoning。
03. Visual Agent Evaluation
This time, besides the model itself, I found that Kimi has also quietly launched a new set of Office Assistant Agents, focused on handling various document needs in daily work such as PPT, Word, and Excel.
Let's see if we can put an end to the era of hand-crafting Office documents?
Document Agent
Prompt: Against the backdrop of the "dual carbon" goals, discuss the practical significance of green and low-carbon development for China's high-quality economic development. Output as a 1500-word PDF file, magazine interior page style, two-column layout.
It seems no further description is needed; the writing style fits the magazine's requirements, and the magazine interior page layout also looks good. Currently, this document assistant seems to have no competitors on the market.
04. Agent Cluster
According to the official introduction, the Agent Cluster is a brand-new exploration of K2.5—evolving from a single Agent into an Agent team.When facing complex tasks, K2.5 can autonomously create its own "clones," forming teams of different roles on demand and working in parallel.。
So I tested this prompt:
Please have five well-known Chinese artists or IP designers with completely different styles design for me 5 series, each with 10 WeChat red envelope cover images themed on the Year of the Horse. I hope the styles are all different.
Kimi K2.5 created 5 designers with different styles and assigned them image generation tasks.
Let's take a look at the generated results together:
Among the 5 styles, the new ink-wash style and the ink-wash style are somewhat similar. Overall quality is quite good, the images are clear, and the patterns look nice. They can be used directly as cover images.
Usually, when looking for past exam papers on my own, it's hard to find them all at once. Even when I finally find a relatively comprehensive collection, downloading requires payment.
The Agent Cluster can also help me package and organize 10 years of past exam papers and explanations, with one-click download. This is truly impressive.
Prompt: Batch download the past 10 years of national civil service exam administrative aptitude test papers and answer explanations.
In addition, data collection and visualization work can be run together.
Prompt: Conduct a detailed investigation of the GDP and fluctuation changes of each province in China over the past 10 years, and visualize them in the form of a China map heatmap. By default, display the 2025 GDP, and support switching years.
For example, Advisor Liang is responsible for collecting data, Descartes is responsible for developing an interactive heatmap webpage, and Jingqiu is responsible for integrating the data collected by Advisor Liang into the webpage.
Then we got a genuinely usable heatmap webpage.
05. Some Sharing
Kimi K2.5 focuses on being an all-around Agent model, has gained visual capabilities, and has updated stronger Office and Coding abilities as well as an Agent Cluster with unlimited clones.
This may mean thatlarge models are moving from "search engines and Q&A assistants" toward "personal AI assistants" and even "commercial digital employees." The future competitive point is not just parameters and leaderboards, but who can complete more steps for you in real work.。
Many repetitive mental tasks will be repriced. Content production, basic development, design drafts, and material organization—these time costs that used to be accumulated through experience will be directly flattened by models.
In the end, what matters is whether your aesthetic sense is up to standard and whether your strategy is effective.
Getting things done under your supervision is the new watershed.
Original link:Kimi launches the new K2.5 model; is domestic AI starting to compete on aesthetics?