Why Smart Developers Use Kimi For Design

Author: Better Stack · Watch on YouTube

Video authored by the channel shown, not by IANIA; all rights belong to the creator. Transcript and summary generated automatically; the original-language content is preserved.

Original language: English

Original video: Why Smart Developers Use Kimi For Design · Channel: Better Stack · 06:11 · Watch on YouTube

Topics: AI evaluation and benchmarks, Software engineering with AI, AI content generation

Summary

The video compares Kimi with Claude and OpenAI across five design and development tasks: a landing page, a 3D scene, a component library, a mobile app, and a game. Kimi wins the web design round, while Astra performs best at 3D modeling, all three models tie on the component library, and none impress on the coffee app. Claude wins the game challenge, leading to the conclusion that each model has different strengths and it can be useful to switch between them. The presenter also notes that running all five Kimi demos cost $6.

Key points

The first challenge was a landing page for a developer-focused Postgres job queue; Kimi was judged the winner for its visuals, animation, typography, and overall balance.
Astra produced the strongest floating-island 3D model, while Kimi’s result was described as blocky and heavily clipped.
The three component libraries were considered roughly equal, so each model received a point.
All three models performed poorly on the coffee-rating mobile app; the presenter said it would likely require many rounds of prompting and refinement.
Claude won the game challenge with stronger animations and more polished details.
The presenter summarizes Kimi as strong at web design, Claude as stronger at interactive demos and games, and OpenAI as strong at 3D modeling.
The presenter says Kimi cost $6 to run all five demos; the cost of Astra and Fireball through the subscription was unclear.

Chapters

Start	Chapter	Summary
00:00	Test setup and landing page	The presenter introduces five challenges and compares Claude, OpenAI, and Kimi on a landing page for a Postgres job queue, with Kimi judged the winner.
02:03	3D floating island	The models create a day-to-night floating island, and Astra is judged best while Kimi’s result has blocky visuals and clipping.
03:05	Component library and mobile app	The component libraries are considered a three-way tie, while all three coffee-rating app designs are criticized as generic and unsuccessful.
04:40	Game challenge and overall results	Claude wins the robot game challenge, and the presenter concludes that each model has different strengths before noting Kimi’s $6 cost for the five demos.

Quotes

personally I think this is the best one. — 01:31

The winner here is definitely Astra. — 02:34

I’m going to give a point to all three here. — 03:37

Claude is much better. OpenAI is definitely the loser here. — 05:41

Kimmy cost just $6 to run all five demos. — 06:11

Concepts and entities

Kimi, Claude, OpenAI, Astra, Fireball, Design Arena, Kimi K3

Full transcript (original language)

00:00 If you head over to design arena, sitting right at the top is not a model by Claude, OpenAI, or Google. It’s Kimi. So today I wanted to test how true this really is. We’ll run through five different demos: web design, 3D modeling, UI components, mobile app design, and game dev. We’ll compare Kimi to models like Astra and Fireball, and I’ve been blown away by some of the results here. I think with a little bit of steering you can now produce production-ready work with these models. But I’ll let you be the judge. We’ll

00:30 score each round fairly and then choose the winner at the end. Okay, so the first task was to design a landing page for a developer-focused product. In this case, it’s a job queue for a Postgres database. The first run here is from Claude. The font choice, a little bit odd, I’m not going to lie. Um the animation here is really good though. You can see items moving between lines in the queue. That was really nice touch. We’ve got the install button here. Overall, the design looks pretty good, although I would say the font is a

01:00 little bit weird. We have nice shadows here, decent examples of the different features of the product. All of these graphics are really well done. Overall, I’d say Claude did a pretty good job there. Let’s look at OpenAI. Slightly better use of fonts, although the graphics don’t look nearly as good. If we scroll back up to the Claude version, you can see we’ve got an animated version here, which looks really nice. OpenAI did static. Overall, the graphics as well don’t look nearly as polished. They look pretty crap, to be honest. Like look at this chart here compared to the charts on

01:31 yeah, on Claude’s side. So you can see Claude did a much better job there. So overall, OpenAI’s attempt not quite as good. Now, if we look at Kimi, personally I think this is the best one. We’ve got the animated graphic on the side, much better use of fonts. This to me already looks like something I could just deploy and I’m I’d be happy with. And I also underlined lives in Postgres here. I love the UX here for for NPM install. The graphics also look really nice, on par with Claude, I’d say. The The whole page looks really balanced, and then finally a call to

02:03 action at the bottom. Hopefully, you would agree. I think Kimmy has definitely won this round. Okay, the next task was to generate a 3D model of a floating island that animates between a day and night cycle. This is Claude’s attempt. So, it looks okay. You can see there’s some slight weird clipping on the waterfall here. If we switch between the day and night cycle, it looks okay. Although, everything does look a little bit blocky. If we look at Open AI, this is from Astra. So much better. You know, look at how sort of weird the trees look on Claude’s version, very blocky.

02:34 Astra did a much better attempt at this. The waterfall is still clipped, although overall looks much nicer. The whole composition looks better to me. We can see if we go between the day and night cycle, the lighting is also is also way less harsh. And then if we look at Kimmy, super blocky here. The waterfall is super clipped. Overall, I’d say this is probably the worst result of the three. So, for 3D design, Kimmy probably doesn’t hold up too well. I’d say the winner here is definitely Astra. And if you’re enjoying this, guys, you’d do me a massive favor by subscribing to the

03:05 channel. Now, let’s get back to the video. The next task was to design a component library. We’ve got Kern from Claude, Former from Open AI, and then Meridian from Kimmy. Now, this one is a really difficult one to judge because I think they all look pretty good. So, we’ve got some button components here. You can see we’ve got inputs, super well balanced, and we’ve got some nice focus elements there. Looks nice. If we look at the version from Open AI, buttons look really nice with the icons. We’ve got some inputs here as well with focus elements. Looks really nice. We’ve got things like checkboxes and radios. And

03:37 then we got the version from Kimmy. Again, buttons look really good. The inputs look super nice. Focus element as well, slightly more of a glow style there. We’ve also got radio buttons, checkboxes. So, overall, I think this is super equal. I’d find it difficult to say which exactly is the best one. Probably comes more down to taste. So, I’m going to give a point to all three here. The next task was to design a mobile app for a coffee rating application. This was the worst round in my opinion. I think every model actually did a pretty poor attempt at this. This is the version from Claude. Looks super

04:08 generic and basic and I would never be happy to release something like this. Very ugly design overall. We’ve got the version from OpenAI. It’s a slight improvement, but again super blocky and generic. I feel like with these sorts of requests, one shotting it is not going to be easily possible. You probably have to refine this over many, many prompts. And again, we’ve got the version from Kimmi. Super similar to the OpenAI version. Blocky, very basic and generic. Overall, I wouldn’t be happy to use any of these. So, so unfortunately, because

04:40 they are all equally terrible, they’re all going to get no points. The final challenge was to generate a game and design all of the assets for it from scratch as well. So, we’ve ended up with this robot jumping game. This is the version from Claude. So, the model looks okay. It’s pretty good detail on the eye. You can see it also blinks there as well. And the animation’s pretty nice. We can also jump. And then we’ve got this flying animation, which looks pretty good. The walking animation also looks really nice and we’ve got the smoke coming out the back. If we look at the version from OpenAI,

05:10 slightly uglier for sure. The animation doesn’t work at all. It almost runs backwards. The flame also looks way more basic. And you can see we’ve got this animation of a some smoke maybe shooting off the screen there. Looks overall pretty weird. I’d say this is a pretty poor result. And finally, we’ve got the version from Kimmi. Overall, I’d say it looks slightly better than the OpenAI version. The feet actually move in the correct direction. The eye looks okay. Actually follows the mouse. Slightly less polished than the Claude version. The flame looks okay. Much better than the

05:41 OpenAI version. I’d say overall, Claude is much better. OpenAI is definitely the loser here. And then Kimmy comes in the middle. So, the point for this round is going to go to Claude. So, overall, it seems like each of the models definitely has its strengths and weaknesses. Kimmy seems very good at things like web design. Claude is much better at interactive demos and games. And then Open AI really shines with 3D modeling. So, depending on your requirements, it’s probably best actually to switch between each of these three models to get the

06:11 best results. Now, I ran Astra and Fireball through the subscription, so I wasn’t sure exactly how much that cost. But, Kimmy cost just $6 to run all five demos. And if you want to see how well Kimmy K3 performs at other software engineering tasks, we’ve got a video for that right here.