A side-by-side comparison of four large-scale models: Kimi K2, DeepSeek, Grok 4, and Claude 4. Which one is the true champion?

A few days ago, the Dark Side of the Moon suddenly dropped a bombshell: the Kimi K2 was officially released. Total specifications: 1T, activation parameters: 32B, directly...open source, canfreeCommercial.

The parameters surged to the trillion level in one go, breaking the current record.open sourceThe model's record.

Among a host of authoritative benchmarks, K2 achieved [a certain result].open sourceThe model's best performance, especially in programming, tool usage, and mathematical reasoning, is very strong.

The new model K2 is now used by default on the Kimi website.

The API is also launched simultaneously, supporting 128K long contexts. The pricing is low: 4 yuan per million tokens as input, and 16 yuan as output, which is approximately... Claude One-fifth of the cost.

I've also had a hands-on experience with it over the past few days, testing several typical application scenarios, and even getting some... Claude A comparative review was conducted on Opus, DeepSeek R1 0528, and Grok 4.

Forget about rankings and specifications; let's just use it and see how it performs. We'll explain which ones are usable and which are good in detail below.

I selected a few mainstream ones.Large ModelCompared with K2, using the samePrompt wordsCompare them separatelyLarge ModelVisual pages, web design, 3D games, and Chinese writing skills.

(For DeepSeek, I tested DeepSeek R1 0528 and DeepSeek V3 0324 respectively, and the results showed that DeepSeek R1 0528 had better overall performance.)

Case 1 Visual Page

I first tested a standardized, general task—visualizing a data dashboard. The task wasn't complex, but it clearly demonstrated several key capabilities of the model: its ability to understand the user's design intent, the consistency of structure and style between modules, and whether the generated code could be run directly.

Prompt wordsGenerate a web-based data dashboard page containing the following modules:
Sales trend line chart (last 30 days)
Pie chart showing regional distribution (different provinces)
User activity heatmap (daily-hourly distribution)
Top summary metrics (number of users, conversion rate, average daily revenue)
Use Chart.js or Plotly, with a consistent style and a clean, modern card design.

Claude 4. The Opus task is half completed. The style is fine, according to my instructions.Prompt wordsIt's here, but unfortunately several charts are blank. There's an awareness of templates, but the actual content is missing.

Grok 4 has all the features you'd expect, but its art style feels a bit dated, so it's only usable to the point of being unusable.

DeepSeek R1 0528 has a good overall chart and page completion. However, there are some extra...Prompt wordsThe interactive options are not mentioned in the game and are not interactive at all; they are just stickers. The heatmap layout is messed up.

K2's overall style is quite good, with a clean structure and pleasing color scheme. The cards and charts are beautifully designed. Surprisingly, K2 correctly modeled the heatmap, but there is a minor flaw—some data extends beyond the card's dimensions.

In the past, when it came to coding assignments, even if there were long queues and fees, many people would still prioritize overseas options.Large ModelAfter all, stability and ease of use have long been the default options;

But this case presented an unusual situation—the domestically produced model performed more stably and better met the requirements. This was somewhat unexpected.

Case 2 Alarm Clock

This need is actually quite common, and many people are using it for the first time. AI When starting a project, we always begin by "replicating the Pomodoro Technique".

A simple alarm clock webpage can reveal a lot. Does it have animation effects? Is the design aesthetically pleasing? Do the buttons respond when clicked? Is the code robust? It's not about showcasing features, but rather truly testing whether the model "can get things done."

I tried a different style this time and ran a version.

Prompt wordsPlease design a pop art style alarm clock webpage interface, including: current time display (large text), alarm clock function, and timer function.

Claude 4. Opus has a great aesthetic, a beautiful page layout, and many detailed features, including alarm clock labels and interactive effects between cards and buttons. Unfortunately, it doesn't display the real-time time, alarm settings don't respond when entered, and the timer function is unusable.

Grok 4 uses a dynamic gradient background, relying entirely on Comic Sans and contrasting colors to maintain the pop art aesthetic. The interaction is rather rough, lacking state-based interaction, but the functionality is complete and meets a usable standard.

The DeepSeek R1 0528's Pop Art style is absolutely top-notch, with a clever combination of color scheme and text shadows. The buttons have status-linked functionality, but there's no real-time time display, and the alarm and timer functions are unusable.

The K2's interface has a comic book feel, with eye-catching color schemes and animations. The time display moves rhythmically, almost like it's breathing, and the buttons are also status-responsive. Functionally, both the alarm clock and timer work perfectly.

Overall, both Grok 4 and K2 generated "usable" web pages on the first try. However, K2's pages are more aesthetically pleasing and have better animations.

Case 3: 3D Shooting Game

This task tests whether the model can comprehensively utilize Three.js to build a complete 3D game scene, realize view control, interactive shooting, particle effects and interface logic, and generate a working web prototype.

Prompt wordsPlease use Three.js to create a first-person perspective.Simple 3D space shooting game, gameplay is as follows:
Players can freely move their camera and click to fire lasers against a starry sky background, scoring points for hitting floating asteroids.
Add particle explosion effect upon impact
Displays a scoreboard + time countdown + victory notification.
BoxGeometry can be used to fake asteroids, but the page must be functional.

Claude The code generated by Opus and Grok 4 won't even run. I sent this code to others...Large ModelPlease help me analyze this.Claude 4. The code generated by Opus suffers from a triple problem: "render not initialized + view logic error + core logic missing," essentially meaning it was submitted halfway through writing.

The issue with Grok 4 was a deployment mismatch. After modifying the deployment-related code, it could indeed be opened.

The background design is a starry sky, which is quite immersive. However, the asteroids are all gray squares with no crosshair, making them impossible to shoot. The page is viewable, but the content is purely decorative.

DeepSeek R1 0528 added a game start screen, and the shooting function also works. The 3D effect of the small squares on the screen is quite good, but the camera angle is uncontrollable, making the game unplayable. After running for a few seconds, the screen freezes completely.

K2 is the only one that's actually playable. It has a starry background, asteroids of varying sizes, a crosshair, and shooting mechanics—basically all the basic functions are there. The UI is complete, and it even thoughtfully includes gameplay hints so you don't get lost when you first open it. This goes beyond just "being able to write code."SimpleIt's more like a model output with product awareness.

CESE 4 Chinese Writing

Copywriting is also a common scenario. This time, I directly asked four people to write copy.Large ModelLet's write a novel together and see who writes the more moving one.

Prompt wordsPlease write a fan fiction novel featuring Sun Wukong and Lin Daiyu as the main characters, blending classical elements with fantasy elements, with a word count of over 3000 words.

Grok 4 is about 4,200 words long. The content is rather formulaic, the plot feels a bit like a diary, and there is very little description of emotional ups and downs, so it doesn't feel very relatable.

Claude 4. Opus wrote approximately 3200 words. The chapter-based format is clear, and the overall structure of the article is good. The language level is average.

DeepSeek R1 0528 quickly reached 6800 words; the writing is excellent, giving the feeling of reading a professional long novel. The background setting and character descriptions are very well done.

Kimi K2 wrote 5100 words. Kimi first planned the overall content and plot. The plot wasn't particularly outstanding, but the writing style was still quite good.

After testing these items, I feel the difference is still quite obvious.

Some models are ready to use right out of the box, while others are still just "looks similar." Issues like pages not running, logical gaps, and chaotic structures can be easily detected through testing.

Coincidentally, just hours after the K2 was released, Sam Altman postponed his planned release this week.open sourceof GPT-4o-mini "Let's polish the interior a bit more."

But after seeing the K2's performance, and then seeing this sudden braking maneuver, everyone understands in their hearts—for the first time, domestically produced models have truly come close to "..."open sourceThe home ground of the "Great War".

The competition continues, but the game has changed.